Topological Duality for Distributive Lattices: Theory and Applications (Cambridge Tracts in Theoretical Computer Science, Series Number 61) [1 ed.] 1009349694, 9781009349697

Citation preview

Cambridge Tracts in Theoretical Computer Science 61 Topological Duality for Distributive Lattices Introducing Stone–Priestley duality theory and its applications to logic and theoretical computer science, this book equips graduate students and researchers with the theoretical background necessary for reading and understanding current research in the area. After giving a thorough introduction to the algebraic, topological, logical, and categorical aspects of the theory, the book covers two advanced applications in computer science, namely in domain theory and automata theory. These topics are at the forefront of active research seeking to unify semantic methods with more algorithmic topics in finite model theory. Frequent exercises punctuate the text, with hints and references provided. M A I G E H R K E is directeur de recherches in computer science at the French National Centre for Scientific Research (CNRS) working at the department of mathematics of Université Côte d’Azur in Nice. Her main contributions are in Stone duality, canonical extensions, and applications in logic and theoretical computer science. S A M VA N G O O L is maître de conférences at the Research Institute for Foundations of Computer Science (IRIF) at Université Paris Cité. His main contributions are in duality theory and logic in mathematics and computer science.

Published online by Cambridge University Press

CAMBRIDGE TRACTS IN THEORETICAL COMPUTER SCIENCE 61 Editorial Board S. Abramsky, Department of Computer Science, University of Oxford P. H. Aczel, School of Computer Science, University of Manchester Y. Gurevich, Microsoft Research J. V. Tucker, Department of Computer Science, Swansea University Titles in the series A complete list of books in the series can be found at www.cambridge.org/computer-science. Recent titles include the following: 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54. 55. 56. 57. 58. 59. 60.

T. F. Melham Higher Order Logic and Hardware Verification R. Carpenter The Logic of Typed Feature Structures E. G. Manes Predicate Transformer Semantics F. Nielson & H. R. Nielson Two-Level Functional Languages L. M. G. Feijs & H. B. M. Jonkers Formal Specification and Design S. Mauw & G. J. Veltink (eds) Algebraic Specification of Communication Protocols V. Stavridou Formal Methods in Circuit Design N. Shankar Metamathematics, Machines and Gödel’s Proof J. B. Paris The Uncertain Reasoner’s Companion J. Desel & J. Esparza Free Choice Petri Nets J.-J. Ch. Meyer & W. van der Hoek Epistemic Logic for AI and Computer Science J. R. Hindley Basic Simple Type Theory A. S. Troelstra & H. Schwichtenberg Basic Proof Theory J. Barwise & J. Seligman Information Flow A. Asperti & S. Guerrini The Optimal Implementation of Functional Programming Languages R. M. Amadio & P.-L. Curien Domains and Lambda-Calculi W.-P. de Roever & K. Engelhardt Data Refinement H. Kleine Büning & T. Lettmann Propositional Logic L. Novak & A. Gibbons Hybrid Graph Theory and Network Analysis J. C. M. Baeten, T. Basten & M. A. Reniers Process Algebra: Equational Theories of Communicating Processes H. Simmons Derivation and Computation D. Sangiorgi & J. Rutten (eds) Advanced Topics in Bisimulation and Coinduction P. Blackburn, M. de Rijke & Y. Venema Modal Logic W.-P. de Roever et al. Concurrency Verification Terese Term Rewriting Systems A. Bundy et al. Rippling: Meta-Level Guidance for Mathematical Reasoning A. M. Pitts Nominal Sets S. Demri, V. Goranko & M. Lange Temporal Logics in Computer Science B. Jacobs Introduction to Coalgebra S. Mihov & K. U. Schulz Finite-State Technology

Published online by Cambridge University Press

Topological Duality for Distributive Lattices Theory and Applications MAI GEHRKE CNRS and Université Côte d’Azur S A M VA N G O O L Université Paris Cité

Published online by Cambridge University Press

Shaftesbury Road, Cambridge CB2 8EA, United Kingdom One Liberty Plaza, 20th Floor, New York, NY 10006, USA 477 Williamstown Road, Port Melbourne, VIC 3207, Australia 314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India 103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467 Cambridge University Press is part of Cambridge University Press & Assessment, a department of the University of Cambridge We share the University’s mission to contribute to society through the pursuit of education, learning and research at the highest international levels of excellence. www.cambridge.org Information on this title: www.cambridge.org/9781009349697 DOI: 10.1017/9781009349680 © Mai Gehrke and Sam van Gool 2024 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press & Assessment. First published 2024 Printed in the United Kingdom by CPI Group Ltd, Croydon CR0 4YY A catalogue record for this publication is available from the British Library A Cataloging-in-Publication data record for this book is available from the Library of Congress ISBN 978-1-009-34969-7 Hardback Cambridge University Press & Assessment has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Published online by Cambridge University Press

Contents

Preface Acknowledgments

page vii xv

1

Order and Lattices 1.1 Preorders, Partial Orders, and Suprema and Infima 1.2 Lattices 1.3 Duality for Finite Distributive Lattices

1 1 9 18

2

Topology and Order 2.1 Topological Spaces 2.2 Topology and Order 2.3 Compact Ordered Spaces

29 29 37 42

3

Priestley Duality 3.1 Prime Filters and Ideals 3.2 Priestley Duality 3.3 Boolean Envelopes and Boolean Duality

51 51 61 76

4

Duality Methods 4.1 Free Distributive Lattices 4.2 Quotients and Subs 4.3 Unary Operators 4.4 Modal Algebras and Kripke Completeness 4.5 Operators of Implication Type 4.6 Heyting Algebras and Esakia Duality 4.7 Boolean Closure and Alternating Chains

86 87 96 112 121 130 139 149

5

Categorical Duality 5.1 Definitions and Examples of Categories 5.2 Constructions on Categories 5.3 Constructions inside Categories

154 154 158 173

Published online by Cambridge University Press

vi

Contents

5.4

Priestley Duality Categorically

188

6

Omega-Point Duality 6.1 Spectral Spaces and Stone Duality 6.2 The Omega-Point Adjunction 6.3 The Omega-Point Duality 6.4 Duality for Spaces of Relations and Functions

196 196 200 205 211

7

Domain Theory 7.1 Domains and Hoffmann–Lawson Duality 7.2 Dcpos and Domains That Are Spectral 7.3 Bifinite Domains 7.4 Domain Theory in Logical Form

227 230 246 256 265

8

Automata Theory 8.1 The Syntactic Monoid as a Dual Space 8.2 Regular Languages and Free Profinite Monoids 8.3 Equations, Subalgebras, and Profinite Monoids 8.4 Open Multiplication

273 274 292 309 325

Bibliography Notation Index

330 339 346

Published online by Cambridge University Press

Preface

This book is a course on Stone–Priestley duality theory, with applications to logic and the foundations of computer science. Our target audience includes both graduate students and researchers in mathematics and computer science. The main aim of the book is to equip the reader with the theoretical background necessary for reading and understanding current research in duality and its applications. We aim to be didactic rather than exhaustive; however, we do give technical details whenever they are necessary for understanding what the field is about. Distributive lattice structures are fundamental to logic, and thus appear throughout mathematics and computer science. The reason for this is that the notion of a distributive lattice is extremely basic: It captures a language containing as its only primitives the logical operators “or,” “and,” “true,” and “false.” Distributive lattices are to the study of logic what rings and vector spaces are to the study of classical algebra. Moreover, distributive lattices also appear in, for example, ring theory and functional analysis. The mathematical kernel that makes duality theory tick is the fact that the structure of a lattice can be viewed in three equivalent ways. A distributive lattice is all of the following: (a) (b) (c)

a partially ordered set satisfying certain properties regarding upper and lower bounds of finite sets; an algebraic structure with two idempotent monoid operations that interact well with each other; and a basis of open sets for a particular kind of topological or order-topological space.

The first part of the book will define precisely the vague notions in this list (“certain properties,” “interact well,” “a particular kind of”), and will prove that these are indeed three equivalent ways of looking at distributive lattices. The correspondence between algebraic and topological structure in the last two items of the list can be

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

viii

Preface

cast in a precise categorical form, and is then called a dual equivalence or simply duality. This duality identifies an exciting, almost magical, and often highly useful intersection point of algebra and topology. Historically, Stone showed in the 1930s that distributive lattices are in a duality with spectral spaces: a certain class of topological spaces with a non-trivial specialization order, which are also the Zariski spectra of rings. Stone’s duality for distributive lattices is especially well known in the more restrictive setting of Boolean algebras, obtained by adding an operator “not” to the lattice signature, which satisfies the usual rules of logic: de Morgan’s laws and excluded middle. The restriction of Stone’s duality to Boolean algebras shows that they are in a duality with compact Hausdorff zero-dimensional spaces. While the spaces associated to Boolean algebras are better known than the slightly more general ones associated to distributive lattices, the latter are vastly more versatile, having as continuous retracts, among others, all compact Hausdorff spaces, including connected spaces such as the unit interval of the reals. Nevertheless, Stone’s duality for distributive lattices was for at least 30 years seen by many as a lesser sibling of his duality for Boolean algebras, at least partly due to the fact that the spaces that figure in it are not Hausdorff, and the appropriate functions between the spaces are not all the continuous ones. Priestley’s seminal work in the 1970s lifted this obstacle, by giving a first-class role to the specialization order that figures in Stone’s spectral spaces. Priestley reframed Stone’s duality as one between distributive lattices and certain partially ordered topological spaces, now called Priestley spaces. The first goal in this book is to build up the necessary mathematics to prove Priestley’s duality theorem, which we do in Chapter 3; we also show there how it easily specializes to the case of Boolean algebras. Building up to this first main result, Chapters 1 and 2 will teach the foundations of order theory and topology that we rely on in the rest of the book. A unique feature of this book is that, in addition to developing general duality theory for distributive lattices, we also show how it applies in a number of areas within the foundations of computer science, namely, modal and intuitionistic logics, domain theory, and automata theory. The use of duality theory in these areas brings to the forefront how much their underlying mathematical theories have in common. It also prompts us to upgrade our treatment of duality theory with various enhancements that are now commonly used in state-of-the-art research in the field. Most of these enhancements make use of operators on a distributive lattice: maps between lattices that only preserve part of the lattice structure. The simplest kind of operator is a map between lattices that respects the structure of “and” and “true,” but not necessarily “or” and “false.” If this notion is understood as analogous to a linear mapping in linear algebra, then it is natural to also consider more general binary, ternary, and n-ary operators on lattices, which respect the

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

Preface

ix

structure of “and” and “true” in each coordinate, as long as the remaining coordinates are fixed. The theory of lattices with operators, and dualities for them, was developed in the second half of the twentieth century, roughly in two main chunks. First, in the 1950s, by Jónsson and Tarski, in the case of Boolean algebras, with immediate applications to relation algebra, and the same theory was used heavily a little later and very successfully for modal logic in the form of Kripke’s semantics. However, until the end of the 1980s, the duality theory for distributive lattices with additional operations developed in the form of a great number of isolated caseby-case studies. Starting with the work of Goldblatt, and also of Jónsson and this book’s first author, the general theory of distributive lattices with additional operations came into a mature, more usable, form by the 1990s. This theory is developed in Chapter 4, which also contains the first applications of duality theory, to free distributive lattices, quotients and subspaces, implication-type operators, Heyting algebras, and Boolean envelopes. In the development of the first four chapters of this book, we keep the use of category theory to a minimum. In Chapter 5, we then set the results of the earlier chapters in the more abstract and general framework of category theory. This development then also allows us to show how Priestley’s duality fits well in a more general framework for the interaction of topology and order, which had been developed by Nachbin shortly before. In Chapter 6, we show how the various classes of topological spaces with and without order, introduced by Stone, Priestley, and others, all relate to each other, and how they are in duality with distributive lattices and their infinitary variant, frames. Chapters 7 and 8 contain two more modern applications of duality theory to theoretical computer science, namely to domain theory and to automata theory, respectively. The domain theory that we develop in Chapter 7 is organized around three separate results: Hoffmann–Lawson duality; the characterization of those dcpos and domains, respectively, that fall under Stone duality; and Abramsky’s celebrated 1991 Domain Theory in Logical Form paper. The duality-theoretic approach to automata theory that we develop in Chapter 8 originates in work linking profinite methods in automata theory with duality theory (Gehrke et al., 2008). It is organized around a number of related results, namely: finite syntactic monoids can be seen as dual spaces, and the ensuing effectivity of this powerful invariant for regular languages; the free profinite monoid is the dual of the Boolean algebra of regular languages expanded with residuation operations and, more generally, topological algebras on Boolean spaces are duals of certain Boolean algebras extended by residual operations. As an extended application example, we use duality to give a profinite equational characterization for the class of piecewise testable languages; and we end by discussing a characterization of those profinite monoids for which the multiplication is open.

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

x

Preface

How to Use This Book This is a textbook on distributive lattices, spectral spaces, and Stone and Priestley dualities as they have developed and are applied in various areas at the intersection of algebra, logic, and theoretical computer science. Our aim is to get in a fairly full palette of duality tools as directly and quickly as possible, then to illustrate and further elaborate these tools within the setting of three emblematic applications: semantics of propositional logics; domain theory in logical form; and the theory of profinite monoids for the study of regular languages and automata. The text is based on lecture notes from a 50-hour course in the Master Logique Mathématique et Fondements de l’Informatique at Université Paris 7, which ran in the winters of 2013 and 2014. The fact that it is based on notes from a course means that it reaches its goals while staying as brief and to the point as possible. The other consequence of its origin is that, while it is fully a mathematics course, the applications we aim at are in theoretical computer science. The text has been expanded a bit beyond what was actually said in the course, reaching research monograph level by the very end of the last two, Chapters 7 and 8. Nevertheless, we have focused on keeping the spirit of a lean and lively textbook throughout, including only what we need for the applications, and often deferring more advanced general theory to the application chapter where it becomes useful and relevant. While the original course on which the book is based covered the majority of all the chapters of the book, there are several other options for its use. In particular, a basic undergraduate course on lattices and duality could treat just Chapters 1 through 3 and possibly selected parts of 4, 5, and/or 6. The applications in the second part are fairly independent and can be included as wanted, although the domain theory material in Chapter 7 requires at least skeletal versions of Chapter 5, and Chapter 6 in its entirety. The first part, Chapters 1 through 6, is a graduate-level “crash course” in duality theory as it is practiced now. Chapter 1 introduces orders and lattices, and in particular the distributive lattices that we will be concentrating on, as well as the finite case of Stone duality, where topology is not yet needed. Chapter 2 introduces the topological side of the dualities. In this chapter, we elaborate the interaction between order and topology, which is so central to the study of spaces coming from algebraic structures. For this purpose we have bent our philosophy of minimum content somewhat by introducing the class of stably compact spaces and Nachbin’s equivalent class of compact ordered spaces. We believe that this setting provides the right level of generality for understanding the connection between Stone’s original duality for spectral spaces and Priestley duality. The class of stably compact spaces, being the closure of spectral spaces under continuous retracts, is also a more robust setting than spectral spaces for a number of further applications that we do not cover

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

Preface

xi

in this book, such as continuous domain theory and duality for sheaf representations of algebras. The basic mathematical content of Priestley duality is given in Chapter 3. Chapter 4 introduces the most important general methods of modern duality theory: duality for additional operations and sub-quotient duality, which then allows us to immediately give first applications to propositional logics. Chapter 5 then introduces categorical concepts such as adjunctions, dualities, filtered colimits, and cofiltered limits, which play a fundamental role in duality theory. This allows us to give a full categorical account of Priestley duality by the end of the chapter. Chapter 6 treats the Omega-Point duality and Stone’s original duality for distributive lattices and makes the relationship between these dualities and Priestley’s version clear. The duality theory developed in the first six chapters of the book is applied to two different parts of theoretical computer science in the last two chapters, which provide an entry into research-level material on these topics. These two chapters are independent from each other, and have indeed traditionally been somewhat separate in the literature, but our treatment here shows how both topics in fact can be understood using the same duality-theoretic techniques that we develop in the first part of the book. When using this book for a course, a lecturer can freely choose material from either or both of these chapters, according to interest. Chapter 7, on domain theory, contains a duality-theoretic exposition of the solutions to domain equations, a classical result in the semantics of programming languages. Chapter 8 develops a duality theory for algebraic automata theory, and shows in particular how finite and profinite monoids can be viewed as instances of the dual spaces of lattices with operators that we study in this book. We have given some bibliographic references throughout the text. We want to emphasize here that these references are not in any way meant to give an exhaustive bibliography for the vast amount of existing research in duality theory. They are rather intended as useful entry points into the research literature appropriate for someone learning this material, who will then find many further references there. Furthermore, at the end of several chapters, we have added a small number of additional notes giving technical pointers pertaining to specific topics discussed there – again, we do not mean to imply exhaustivity. When we introduce special or less standard notation, we use “Notation” blocks, which are occasionally numbered when we need to refer back to them later. The book ends with a listing of the most-used notations and an index of concepts. In each chapter, all numbered items follow one and the same counter, with the exception of exercises, which follow a separate numbering, indicating not only chapter but also section. On the topic of exercises: this book contains many of them, varying greatly in difficulty. In earlier chapters, many of the exercises are routine verifications, but necessary practice for a learner who wants to master the material.

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

xii

Preface

Especially in later chapters, there are exercises that could be viewed as small research projects, although we refrain from stating open problems as exercises: for the less obvious exercises, we have included hints and references where available. Comparison to Existing Literature and Innovative Aspects The first part of this book, Chapters 1 through 6, covers quite classical material and may be compared to existing textbooks. The closest are probably Distributive Lattices (Balbes and Dwinger, 1975) and Introduction to Lattices and Order (Davey and Priestley, 2002). Another classical gentle introduction to the field, but focusing more on point-free topologies and frame theory than we do here, is Topology via Logic (Vickers, 1989). Of these, Balbes and Dwinger (1975) is probably the closest in spirit to our treatment, as it gets to the duality quickly and then applies it. However, that book’s applications to algebras of propositional logic focus on varieties that are less central today. Davey and Priestley’s textbook has been very successful and has in particular managed to attract a theoretical computer science readership to these topics. However, it focuses more on the lattices and order per se and the duality is covered only as one of the final crowning chapters. Davey and Priestley’s book is therefore an excellent way in to ours, and we recommend it as supplemental reading in case students need additional details or to build up mathematical maturity. The textbooks Introduction to Boolean Algebras (Givant and Halmos, 2008) and Duality Theories for Boolean Algebras with Operators (Givant, 2014) are also relevant but are of course restricted to the Boolean setting. There are also a number of classical references in lattice theory by Grätzer, the most recent versions being Lattice Theory: Foundation (Grätzer, 2011) and General Lattice Theory (Grätzer, 2003), both of which contain material on duality theory and its applications to lattice theory. Here we aim to get the dualities in place as soon as possible and then use them. Where we differ the most from the existing books within this first part is with our emphasis on the interaction between order and topology in Chapter 2, and in placing Priestley duality within the wider context of category theory (Chapter 5) and Omegapoint duality (Chapter 6). Chapter 2 provides a textbook-level didactic account of the interaction between topology and order culminating with the equivalence between Nachbin’s compact-ordered spaces and stably compact spaces. In Chapter 4 we develop duality theory methods for analyzing the structure of distributive lattices and operators on them. All of these topics have become central in research in recent decades but are so far difficult to access without delving in to the specialized literature. Perhaps the most important omission of this book is the theory of canonical extensions, which was central to the previously mentioned work of Jónsson and

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

Preface

xiii

others, in addition to duality. While this theory is very close to the hearts of both authors of this book, and closely related to duality theory, this book is not about that, and canonical extensions thus do not play a big role in this book, at least not explicitly. Still, we will occasionally make reference to canonical extensions where appropriate. Along with and closely related to this omission, we decided to take a point-set rather than a point-free approach to the topics of this book. Point-free approaches focus on the algebraic side of duality and thus avoid the point-set world of topology, which inherently involves non-constructive principles. Duality is in a sense the justification for the point-free approach since it makes the link between the algebraic and the point-set worlds. In this book we remain fully anchored in the set-theoretic approach to topology, in particular making use of the axiom of choice as necessary. We do this as it is more easily accessible for a general audience, and because our end applications in denotational semantics and profinite algebras in automata theory are, in their currently practiced form, focused on point-set topology. That being said, our focus on duality shows the way and familiarizes the reader with the dual, point-free approach, thus making them ready to embrace this approach. In this direction, one of our hopes with this book is that it will entice some readers to learn about canonical extensions and related point-free techniques. We believe the technique of canonical extensions to be complementary to, and at least as important as, duality, but so far less well established in the literature. Many research monographs include similar material to that of the first part of this book, but are not explicitly targeted at readers who are first learning about the field, while this is a primary aim of our book. Classical such monographs, closest in content to the first part of this book, are Stone Spaces (Johnstone, 1986) and A Compendium of Continuous Lattices (Gierz et al., 1980), re-edited as Continuous Lattices and Domains (Gierz et al., 2003). More recently, the monograph Spectral Spaces (Dickmann et al., 2019) studies the same class of spaces as we do in this book, but comes from a ring-theoretic perspective and emphasizes less the ordertheoretic aspects. The monograph Non-Hausdorff Topology and Domain Theory (Goubault-Larrecq, 2013) is close in spirit to our treatment in Chapter 2, especially in its treatment of stably compact spaces, and also addresses a theoretical computer science audience. A difference with our treatment here is that Goubault-Larrecq (2013) is focused on non-Hausdorff topologies and therefore does not treat the (Hausdorff) patch topology as central, as we do here. Related to our Chapter 6 is the monograph Frames and Locales (Picado and Pultr, 2012) focused on frames and point-free topology, and Chapter 6 of this book can be used as a preparation for jumping into that work. The applications to domain theory and automata theory are treated in Chapters 7 and 8, respectively. These two applications, and in particular the fact that we treat them in one place, as applications of a common theory, are perhaps the most

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

xiv

Preface

innovative and special aspects of this book. Domain theory is the most celebrated application of duality in theoretical computer science and our treatment is entirely new. Automata theory is a relatively new application area for duality theory and has never been presented in textbook format before. More importantly, both topics are at the forefront of active research seeking to unify semantic methods with more algorithmic topics in finite model theory. While previous treatments remained focused on the point of view of domains/profinite algebra, with duality theory staying peripheral, a shared innovative aspect of the presentations of these topics in this book is that both are presented squarely as applications of duality. Finally, a completely original contribution of this book, which emerged during its writing, precisely thanks to our treatment of the two topics as an application of a common theory, is the fact that a notion we call “preserving joins at primes” turns out to be central in both the chapter on domain theory and in that on automata theory. This notion was introduced in the context of automata theory and topological algebra in Gehrke (2016); its application to domain theory is new to this book and reflects a key insight of Abramsky’s Domain Theory in Logical Form. We believe this point to be an exciting new direction for future research in the field that we hope some readers of this book will be inspired to take up.

https://doi.org/10.1017/9781009349680.001 Published online by Cambridge University Press

Acknowledgments

We thank Cambridge University Press, in particular David Tranah and Anna Scriven, for their trust and help with publishing this book. We thank Angela Valente for her very careful copy-editing job, and Vinithan Sethumadhavan for managing the production part of the publication process. We would also like to thank all our colleagues and students who have inspired and encouraged us to complete this book. More specifically, many have used and commented on early versions and their comments have been immensely useful. Samson Abramsky, Jorge Almeida, Achim Jung, Jean Goubault-Larrecq, JeanÉric Pin, and Benjamin Steinberg have all provided us with invaluable advice and support on several of the later parts of the book; their reading of parts of our manuscript at various stages and their very detailed and useful feedback have been of great help. We would also like to thank Jim de Groot and Luca Reggio, who both proofread the manuscript very carefully, and also Célia Borlido and Victor Iwaniack who made many useful comments. Jérémie Marquès has made vital mathematical contributions to the book, also acknowledged specifically in Chapter 8. We are very grateful for the friendship, comments, and support that we have received from those already mentioned above, as well as from many other colleagues and co-authors with whom we have discussed and worked on several of the topics discussed in this book, specifically Clemens Berger, Guram Bezhanishvili, Nick Bezhanishvili, Mikołaj Bojańczyk, Thomas Colcombet, Dion Coumans, Mirna Dˇaamonja, Marcel Erné, Wesley Fussner, Silvio Ghilardi, Charles Grellois, Serge Grigorieff, John Harding, Tomáš Jakl, André Joyal, Alexander Kurz, Vincenzo Marra, Paul-André Melliès, George Metcalfe, Alessandra Palmigiano, Daniela Petrişan, Hilary Priestley, Carlos Simpson, Howard Straubing, Yde Venema, and Glynn Winskel. We are also very grateful to the many students who attended our lectures and worked with the material in this book. This book is primarily written for them.

Published online by Cambridge University Press

xvi

Acknowledgments

We dedicate this book to the memory of Bernhard Banaschewski, Bjarni Jónsson, and Klaus Keimel. Finally, we thank our families for their love and support.

Published online by Cambridge University Press

1 Order and Lattices

In this chapter, we first introduce basic notions from order theory: preorders, partial orders, and lattices. We then zoom in on distributive lattices. In the finite case, we prove from first principles a duality theorem, which is a blueprint for the more advanced duality theorems that follow later in this text. 1.1 Preorders, Partial Orders, and Suprema and Infima A binary relation  on a set P is called • • • • •

reflexive if p  p for all p ∈ P, transitive if p  q  r implies p  r for all p, q, r ∈ P, anti-symmetric if p  q and q  p imply p = q for all p, q ∈ P, a preorder if it is reflexive and transitive, and a partial order if it is reflexive, transitive, and anti-symmetric.

A preordered set is a tuple (P, ) with  a preorder on the set P. A poset (short for partially ordered set) is a pair (P, ≤) with ≤ a partial order on the set P. Two elements p and q are comparable in a preorder  if at least one of p  q and q  p holds, and incomparable otherwise. The adjective “partial” in “partial order” refers to the fact that not all elements in a partial order are comparable. A preorder is called total or linear if any two of its elements are comparable. A total order or linear order or chain is a total preorder which is moreover anti-symmetric. A poset is called an anti-chain if no distinct elements are comparable. The strict part of a partial order is the relation < defined by p < q if, and only if, p ≤ q and p  q. Notice that, if we have access to equality, then to specify a partial order ≤, it suffices to specify its strict part 0 such that the interval (x − ε, x + ε) = {y ∈ R | x − ε < y < x + ε} is entirely contained in U. It is not hard to see that a function f : R → R is continuous with respect to this topology if, and only if, it satisfies the usual epsilondelta definition of continuity (see Exercise 2.1.1). A subspace of a topological space (X, τ) is given by a subset Y ⊆ X and is equipped with the subspace topology, defined as τ  Y = {U ∩ Y | U ∈ τ}. A continuous function f : X → Y between topological spaces is said to be an open mapping provided f [U] = { f (x) | x ∈ U} is open in Y for any open U ⊆ X. Similarly, f : X → Y is said to be a closed mapping provided f [C] = { f (x) | x ∈ C} is closed in Y for any closed C ⊆ X. Further, f is said to be an embedding provided it is injective and f −1 : im( f ) → X is also continuous, where we view f [X] = im( f ) as a topological space in the subspace topology. Finally, f is a homeomorphism provided it is a bijection and both f and f −1 are continuous. Since open sets are closed under unions, for any subset S of a topological space (X, τ), there is a largest open set, int(S), that is contained in S. The set int(S) is called the interior of S; in a formula,  int(S) = {U ⊆ X | U ⊆ S and U ∈ τ}. We note that int is the upper adjoint to the inclusion map Ω(X ) → P (X ) (see Exercise 2.1.3). If x ∈ X is a point and S is any subset of X, then S is a neighborhood

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.1 Topological Spaces

31

of x if x ∈ int(S), that is, if there exists an open set U ⊆ S such that x ∈ U. The interior of a set S is sometimes denoted by S ◦ . Symmetrically, since closed sets are closed under arbitrary intersections, any subset S has a closure, cl(S), which is defined as the smallest closed set containing S; in a formula, cl(S) = {C ⊆ X | S ⊆ C and C is τ-closed}. The map cl is the lower adjoint to the inclusion map C(X ) → P (X ). The closure of a set S is sometimes denoted by S. A subset S of X is called dense if cl(S) = X. Notice that arbitrary intersections of topologies on a fixed set X are again topologies. Accordingly, for any collection S of subsets of X, there is a least topology containing S. We call this the topology generated by S and denote it by S. When τ is a topology on a set X, a collection S of subsets of X is called a subbase for the topology τ if τ = S. Note that S is the collection of subsets of X that can be written as arbitrary unions of finite intersections of elements of S. We say that a subbase S for a topology τ is a base if, for every x ∈ X and U ∈ τ with x ∈ U, there is V ∈ S with x ∈ V ⊆ U. Notice that, in this case, finite intersections are not needed to generate τ, that is, every open of τ is simply a union of elements from S. An equivalent way of saying this is that a subbase B is a base for the topology generated by B if and only if, for every x ∈ X, Bx = {V ∈ B | x ∈ V } is filtering. In particular, the closure of a subbase S under finite intersections is always a base, but bases are more general than that; see Exercise 2.1.5. Let (Xi )i ∈ I be a collection of topological spaces indexed by a set I. The product  space i ∈ I Xi is the Cartesian product of the sets Xi , equipped with the topology generated by the subbase consisting of the sets πi−1 (V ), where i ∈ I and V ⊆ Xi is open in Xi,

 th where πi : j ∈I X j → Xi is the projection onto the i coordinate for i ∈ I. Let X be a topological space. We recall the five main separation axioms that may hold for X. • X is T0 (or Kolmogorov) provided, for all x, y ∈ X with x  y, there is an open U ⊆ X which contains exactly one of x and y. • X is T1 (or Fréchet) provided, for all x, y ∈ X with x  y, there is an open U ⊆ X with x ∈ U and y  U. • X is T2 (or Hausdorff) provided, for all x, y ∈ X with x  y there are opens U, V ⊆ X with x ∈ U and y ∈ V and U ∩ V = ∅. • X is T3 (or regular) provided X is T1 and, for all x ∈ X and closed C ⊆ X with x  C, there are opens U, V ⊆ X with x ∈ U and C ⊆ V and U ∩ V = ∅.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

32

2 Topology and Order

• X is T4 (or normal) provided X is T1 and, for all closed C, D ⊆ X such that C ∩ D = ∅, there are opens U, V ⊆ X with C ⊆ U and D ⊆ V and U ∩ V = ∅. Hausdorff spaces may also be characterized as those spaces satisfying a closed graph theorem, that is, a space X is Hausdorff if, and only if, the graph of any continuous function Y → X is closed (see Exercise 2.1.12). Compactness Let S ⊆ X, where X is a topological space. An open cover U of S is a collection

of open sets U ⊆ Ω(X ) such that S ⊆ U . A subset K ⊆ X is compact provided every open cover U of K contains a finite subcover, that is, a finite subset U  ⊆ U which is also a cover of K. An equivalent definition of compactness using closed sets instead of open sets is the following. For a collection A ⊆ C(X ) of closed sets of X, say that A has the finite intersection property with respect to K if for every finite    subcollection A  of A, A ∩K  ∅. Then, K is compact if, and only if, for every collection A ⊆ C(X ) that has the finite intersection property with respect to K, we   have A ∩ K  ∅. In particular, the space X itself is compact if every open cover of X contains a finite subcover, or equivalently, if every collection of closed sets with the finite intersection property (with respect to X) has a non-empty intersection (see Exercise 2.1.6 for further equivalent definitions of compactness for a space). Example 2.4 The subspace [0, 1] = {x ∈ R | 0 ≤ x ≤ 1} of R with the usual topology is compact (see Exercise 2.1.14). A topological space X is locally compact provided that, for each x ∈ X and each U ∈ Ω(X ) with x ∈ U, there are V ∈ Ω(X ) and K ⊆ X compact such that x ∈ V ⊆ K ⊆ U. Note that, if X is Hausdorff, then compactness implies local compactness, while this is not the case in general (see Exercise 2.1.17). Proposition 2.5 Let X and Y be topological spaces and πY : X × Y → Y the projection onto the second coordinate. If X is compact, then πY is a closed mapping. Proof Let C ⊆ X × Y be closed and suppose y  πY [C]. That is, for each x ∈ X, we have (x, y)  C. Thus, as C is closed, for each x ∈ X, there are Ux open in X and Vx open in Y with (x, y) ∈ Ux × Vx and C ∩ (Ux × Vx ) = ∅. Since x ∈ Ux for each x ∈ X, the collection {Ux | x ∈ X } is an open cover of X. Also, as X is compact, there is a finite subset M ⊆ X such that {Ux | x ∈ M }

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.1 Topological Spaces

covers X. Now, letting V = V ∩ πY [C] = ∅.



33

{Vx | x ∈ M } we have y ∈ V , V ⊆ Y open, and 

We now recall a result, which requires a non-constructive principle, and is often useful for proving compactness. Theorem 2.6 (Alexander subbase theorem) Let X be a topological space and S a subbase for the topology on X. If every cover U ⊆ S of X has a finite subcover, then X is compact. Any proof of the Alexander subbase theorem must necessarily use a nonconstructive principle, and we here use Zorn’s lemma, which will also be crucially used in Chapter 3 in the proof of the Stone prime filter-ideal theorem (Theorem 3.10). Note that while the latter has historically been called a “lemma,” we treat it here as a postulate, as is common, since it is equivalent in Zermelo–Fraenkel set theory to the axiom of choice. We have made the choice to freely use non-constructive principles like Zorn’s lemma in this book, although they could sometimes have been avoided. For more information about equivalences between various choice principles, see, for example, the books by Jech (1973) and Howard and Rubin (1998). Lemma 2.7 (Zorn’s lemma) Let S be a non-empty partially ordered set such that if C ⊆ S is totally ordered, then there exists an upper bound c of C in S. Then, S has a maximal element, that is, there exists s ∈ S such that for any s  ∈ S, if s  ≥ s, then s  = s. Proof that Zorn’s lemma implies the Alexander subbase theorem. We ask the reader to fill in the gaps of this proof sketch in Exercise 2.1.10. We prove the contrapositive statement. Suppose that X is not compact. Zorn’s lemma guarantees that there exists a maximal open cover C of X which does not have a finite subcover. The subcollection C ∩ S of C can be shown to still be a cover of X, using the maximality of C. Now C ∩ S is a cover by elements from the subbase, which cannot have a finite subcover.  An earlier, equivalent, variation of Zorn’s lemma, that we will sometimes use, is the following. Lemma 2.8 (Hausdorff maximality principle) For any totally ordered subset C of a partially ordered set P, there is a totally ordered subset C  ⊇ C which is maximal among totally ordered subsets of P, with respect to subset inclusion.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

34

2 Topology and Order

Exercises for Section 2.1 Exercise 2.1.1 Consider R equipped with the collection τ of subsets U ⊆ R with the property that, for each x ∈ U, there exists ε > 0 such that the interval (x − ε, x + ε) is entirely contained in U. (a) (b) (c)

Show that (R, τ) is a topological space. Show that the collection of intervals (r, s), where both r and s are rational, forms a base for the topology τ on R. Show that f : R → R is continuous if, and only if, for every x ∈ R and every ε > 0, there exists δ > 0 such that, for all y ∈ R with |x − y| < δ, we have | f (x) − f (y)| < ε.

Exercise 2.1.2 Give examples of continuous maps which are: (a) (b) (c)

neither open nor closed; open but not closed; and closed but not open.

Exercise 2.1.3 Let X be a topological space. Prove that the interior map int : P (X ) → Ω(X ) is upper adjoint to the inclusion map ι : Ω(X ) → P (X ), and that the closure map cl : P (X ) → C(X ) is lower adjoint to the inclusion ι : C(X ) → P (X ). Exercise 2.1.4 Let X and Y be topological spaces and f : X → Y a continuous injection. (a) (b)

(c) (d)

Show that f need not be an embedding. Show that f is an embedding if, and only if, the co-restriction of f to im( f ) is a homeomorphism. Here, the co-restriction of f : X → Y to a subspace S ⊆ Y , where im( f ) ⊆ S, is the function with domain X, codomain S, and the same action as f . Show that if the continuous map f is a bijection, then it is a homeomorphism if, and only if, it is open and if, and only if, it is closed. Show that if X is a subset of Y and f : X → Y is the inclusion map, then X is a subspace of Y if, and only if, f is an embedding.

Exercise 2.1.5 Let (X, τ) be a topological space and B and S be, respectively, a base and a subbase for τ.

(a) Show that B = { C | C ⊆ B}. (b) Let T be the closure of S under finite intersections. Show that T is a base for τ. (c) Give an example of a base for a topology which is not closed under binary intersections.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.1 Topological Spaces

(d)

35

Let Y be a topological space. Show that f : Y → X is continuous if, and only if, f −1 (U) is open in Y for each U ∈ S.

Exercise 2.1.6 Let X be a topological space. Show that all of the following statements are equivalent:

(a) For all R ⊆ Ω(X ), if R = X, then there exists a finite R  ⊆ R such that

 R = X.

(b) For all directed D ⊆ Ω(X ), if D = X, then X ∈ D.  (c) For all S ⊆ C(X ), if S = ∅, then there exists a finite S  ⊆ S such that   S = ∅.

(d) For all filtering F ⊆ C(X ), if F = ∅, then ∅ ∈ F . 

(e) For all R ⊆ Ω(X ), S ⊆ C(X ), if S ⊆ R, then there exist finite R  ⊆ R 

and finite S  ⊆ S such that S  ⊆ R .  (f) For all directed D ⊆ Ω(X ) and filtering F ⊆ C(X ), if F ⊆ D, then there exist F ∈ F and U ∈ D such that F ⊆ U. Exercise 2.1.7 (a) Let X be a compact topological space and let C ⊆ X be closed. Show that C is compact. (b) Find a compact space X with a compact subset K which is not closed. (c) Show that if a topological space X is Hausdorff and K ⊆ X is compact, then K is closed. Exercise 2.1.8 (a) Show that the finite union of compact sets is compact. (b) Give an example to show that the intersection of two compact sets need not be compact. Exercise 2.1.9 Let X be a compact topological space and let F be a filter of clopen subsets of X. Prove that, for any clopen set K in X, K ∈ F ⇐⇒ F ⊆ K. Hint. Use the characterization of compactness given in Exercise 2.1.6. Exercise 2.1.10 This exercise is based on Engelking (1989, Exercise 3.12.2, p. 221) and asks you to fill in the details of the proof of the Alexander subbase theorem. Let X be a topological space and S a subbase for the topology on X. (a)

Prove that, in the following sub-poset of (P (Ω(X )), ⊆), C := {C ⊆ Ω(X ) : C is a cover of X and C has no finite subcover}, any totally ordered subset (Ci )i ∈I of C has an upper bound in C. Hint. Show

that i ∈I Ci is in C. Suppose X is not compact. Then, by Zorn’s lemma we can pick a maximal element M of the poset C.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

36

(b) (c)

(d)

2 Topology and Order

Prove that, for any open sets U, V , if V ∈ M and U ⊆ V , then U ∈ M. Prove that, for any finite number of open sets U1, . . . , Un , if Ui  M for every n Ui  M. Hint. Use the maximality of M to get finite 1 ≤ i ≤ n, then i=1

n n Fi ∪ { i=1 Ui } subcovers Fi of M ∪ {Ui } for every i, and show that F := i=1 n is then a finite subcover of M ∪ { i=1 Ui }. Conclude that M ∩ S is a cover of X that does not have a finite subcover.

Exercise 2.1.11 Show that T4 implies T3 , which implies T2 , which implies T1 , which implies T0 . Further, show that all these implications are strict. Exercise 2.1.12 (a) Show that a topological space X is Hausdorff if, and only if, the diagonal relation ΔX = {(x, x) | x ∈ X } (b)

is closed in the product topology on X × X. Let f : Y → X be a continuous function and suppose that X is Hausdorff. Prove that the graph of f , that is, the relation {(y, f (y)) | y ∈ Y }

(c)

is closed in the product topology on Y × X. Conclude that a topological space X is Hausdorff if, and only if, the graph of any continuous function Y → X is closed.

Exercise 2.1.13 Let X and Y be topological spaces. (a) (b)

Prove that, for any y0 ∈ Y , the function i : X → X ×Y defined by i(x) := (x, y0 ) is continuous. Prove that the projection maps πX : X ×Y → X and πY : X ×Y → Y are open.

Exercise 2.1.14 Show that the subspace [0, 1] = {x ∈ R | 0 ≤ x ≤ 1} of R with the usual topology is compact. Exercise 2.1.15 Show that any compact Hausdorff space is normal. Hint. Show first that it is regular. Exercise 2.1.16 (Tychonoff’s theorem) Show that the product of compact spaces is again compact. Hint. A proof can be found in any standard reference on topology, for example, Engelking (1989, Theorem 3.2.4, p. 138). Exercise 2.1.17 (a) Show that any compact Hausdorff space is locally compact. Hint. Use the fact that compact Hausdorff spaces are regular, as proved in Exercise 2.1.15. (b) Find a topological space which is compact but not locally compact.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.2 Topology and Order

37

Exercise 2.1.18 (Quotient space) Let X be a topological space and ≡ an equivalence relation on X. The quotient space of X by ≡ is the space based on X/≡ whose open sets are those U ⊆ X/≡ such that  q−1 (U) = {[x]≡ | [x]≡ ∈ U} = {x ∈ X | [x]≡ ∈ U} is open in X, where q : X → X/≡, x → [x]≡ is the canonical quotient map. (a) (b) (c) (d) (e)

(f) (g)

Show that the topology on X/≡ is the finest topology on X/≡ making q : X → X/≡ continuous. Show that X/≡ is a T1 space if, and only if, every equivalence class of ≡ is closed in X. Show that if X/≡ is a Hausdorff space, then ≡ is necessarily closed as a subset of the product space X × X. Show that if the quotient map is open, then X/≡ is Hausdorff if, and only if, ≡ is a closed in X × X. Show that if f : X  Y is a continuous surjection, then f factors through the canonical quotient map q : X → X/ ker( f ) by a unique continuous bijection f˜ : X/ ker( f ) → Y . Give an example in which f˜ is not a homeomorphism. Show that if f is open or closed, then f˜ is a homeomorphism. But show by giving an example that this condition is not necessary. 2.2 Topology and Order

Let (X, τ) be a topological space. The specialization order of τ is the binary relation ≤τ on X defined by x ≤τ y ⇐⇒ for every U ∈ Ω(X ), if x ∈ U, then y ∈ U. This relation ≤τ is clearly reflexive and transitive and thus, for any topological space, the specialization order is a preorder on X. It is not hard to see that it is a partial order if, and only if, X is T0 . Moreover, for any y ∈ X, the principal down-set, ↓y, of y in the specialization order is the closure of the singleton set {y}. In particular, T1 spaces can be characterized as those spaces having a trivial specialization order. The reader is asked to prove these statements in Exercise 2.2.3. In analysis and algebraic topology, the spaces studied are almost always Hausdorff, so in these fields the interaction with order theory is minimal. However, in applications of topology to algebra and logic, almost all spaces are T0 but not T1 , so in these areas, as in theoretical computer science, the interaction of topology and order plays an important role. A subset of a topological space X is said to be saturated provided it is an up-set in the specialization order, or equivalently, provided it is an intersection of opens.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

38

2 Topology and Order

Note that a subset K ⊆ X is compact if, and only if, its saturation ↑K is compact. We denote by K S(X ) the collection of compact-saturated subsets of X (i.e., subsets of X that are both compact and saturated). As we will see later on, beyond the Hausdorff setting, but in the presence of compactness, K S(X ) is in many aspects the right generalization of the closed subsets. The following fact is often useful and illustrates the consequence of compactness in terms of the specialization order. Recall that, for a subset S of a poset, min S is the (possibly empty) set of minimal points of S. Proposition 2.9 Let X be a T0 space and K ⊆ X compact, then K ⊆ ↑ min(K ), where we consider X in its specialization order. In particular, if K is compact saturated, then K = ↑ min(K ). Proof We first show that if D ⊆ K is down-directed in (X, ≤τ ), then there is a lower bound of D in K. Let D be a down-directed set in (X, ≤τ ) and suppose that K contains no lower bounds of D. Then, the directed collection of open subsets {(↓x) c | x ∈ D} is an open cover of K. Therefore, by compactness, it follows that there is x ∈ D with K ⊆ (↓x) c . In particular, x  K and thus D  K. Now let x ∈ K and, by the Hausdorff maximality principle (Lemma 2.8), let C be a maximal chain in K containing x. Then, by the above argument, C has a lower bound x  ∈ K. Now by maximality of C, it follows that x  ∈ min(K ) and thus K ⊆ ↑ min(K ).  The Lattice of All Topologies on a Set Let X be a set. Note that the collection Top(X ) := {τ ∈ P (P (X )) | τ is a topology} is closed under arbitrary intersections and thus (see Exercise 1.2.2 in Chapter 1) it is a complete lattice in the inclusion order. Infima are given by intersections, while suprema are given by the topologies generated by unions. The least topology on X is the indiscrete topology, while the largest is the discrete topology. We will often make use of the binary join of topologies on a given set X. The interaction of compactness and the Hausdorff separation axiom is illuminated by looking at Top(X ) as a complete lattice. Indeed, by inspecting the definitions (see Exercise 2.2.1), note that Top2 (X ) := {τ ∈ Top(X ) | τ is T2 } is an up-set in Top(X ), while TopComp (X ) := {τ ∈ Top(X ) | (X, τ) is compact}

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.2 Topology and Order

39

is a down-set in Top(X ). It follows that the set of compact-Hausdorff topologies on a set X is a convex subset of Top(X ). The following very useful result tells us that it is in fact an anti-chain. Proposition 2.10 Let X be a set and σ and τ topologies on X with σ ⊆ τ. If σ is Hausdorff and τ is compact, then σ = τ. Proof Since Top2 (X ) is an up-set and TopComp (X ) is a down-set, the hypotheses on σ and τ imply that both are simultaneously compact and Hausdorff. Note that in any T1 space, and thus in particular in a Hausdorff space, any set is saturated. Moreover, in compact-Hausdorff spaces, being closed is equivalent to being compact (and saturated), see Exercise 2.2.7. Also, any set which is compact in a bigger topology remains so in the smaller topology. Thus, we have the following sequence of (bi)implications for any subset U ⊆ X: U ∈ τ ⇐⇒ U c ∈ C(X, τ) ⇐⇒ U c ∈ K S(X, τ) =⇒ U c ∈ K S(X, σ) ⇐⇒ U c ∈ C(X, σ) ⇐⇒ U ∈ σ.



Order-Topologies From topology we get order, but it is also possible to go the other way. Especially in computer science applications where a second-order structure such as a topology is difficult to motivate, topologies induced by orders play an important role; see also the applications to domain theory in Chapter 7. As we have seen, for any topological space, the closures of points are equal to their principal down-sets for the specialization order. Thus, if a topological space X has specialization order ≤, then at least each set of the form (↓x) c , for x ∈ X, must be open. We now proceed in the converse direction. Let (P, ≤) be a partially ordered set. We define several topologies on P for which the specialization order coincides with ≤. • The upper topology on P, ι↑ (P), is defined as the least topology in which ↓p is closed for every p ∈ P. That is, the upper topology is given by ι↑ (P) = (↓p) c | p ∈ P. • The Scott topology on P consists of those up-sets which are inaccessible by  directed suprema. That is, an up-set U ⊆ P is Scott open if, and only if, D ∈ U implies U ∩ D  ∅ for all directed subsets D ⊆ P. We denote the Scott topology on P by σ(P).

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

40

2 Topology and Order

• The Alexandrov topology, α(P), on P is the largest topology on P yielding ≤ as its specialization order. That is, α(P) = {U ⊆ P | U is an up-set}.

The upper, the Scott, and the Alexandrov topologies all have the original order ≤ as their specialization order. In fact, if we denote by Top(P, ≤) the complete lattice of topologies on P yielding ≤ as their specialization order, it is not hard to see that this is precisely the closed interval [ι↑ (P), α(P)] in Top(P, ≤). Clearly, there are order-dual definitions for each of these topologies, which have the reverse of ≤ as their specialization order. For example, the lower topology on P is defined by ι↓ (P) = (↑p) c | p ∈ P. The dual Alexandrov topology has all down-sets as open sets. One can also consider the order-dual of the Scott topology but this is not so common, as the motivation for having closed sets which are stable under directed joins comes from a model of computing in which a computation is considered as the directed join of all its partial computations or finite approximations; more on this in Chapter 7. Using the above “one-sided” topologies as building blocks, we now also define a number of “two-sided” topologies on a partially ordered set (P, ≤). These two-sided topologies are all T1 and thus have trivial specialization order. • The interval topology on P is the join of the upper and lower topologies. That is, ι(P) := ι↑ (P) ∨ ι↓ (P). The usual topology on the reals is in fact the interval topology given by the usual order on the reals. • The Lawson topology on P is the join of the Scott and the lower topologies. That is, λ(P) := σ(P) ∨ ι↓ (P). • Note that the join of the Alexandrov and dual Alexandrov topologies is the discrete topology on P.

Exercises for Section 2.2 Exercise 2.2.1 Prove that Top2 (X ) is an up-set in Top(X ), and that TopComp (X ) is a down-set in Top(X ).

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.2 Topology and Order

41

Exercise 2.2.2 Let X be a compact space, Y a Hausdorff space, and f : X → Y a continuous map. (a)

Show that the map f is closed. Hint. Use Proposition 2.10.

(b)

Show that if f is a bijection, then it is a homeomorphism.

(c)

Show that the co-restriction of f to its image X  im( f ), x → f (x) is a quotient map. In particular, as soon as f is surjective, it is a quotient map (see Exercise 2.1.18).

Exercise 2.2.3 Let X be a topological space. (a)

Show that the specialization order on X is a preorder.

(b)

Show that X is T0 if, and only if, the specialization order on X is a partial order.

(c)

Show that X is T1 if, and only if, the specialization order on X is trivial. Recall that a partial order ≤ is called trivial provided that x ≤ y if, and only if, x = y.

(d)

Show that x ≤ y in the specialization order if, and only if, x ∈ {y}. In other words, show that {y} = ↓y.

(e)

Show that a subset S ⊆ X is an intersection of open sets if, and only if, it is an up-set in the specialization order.

Exercise 2.2.4 Let X be a set and ≤ a partial order on X. Show that a topology τ on X has ≤ as its specialization order if, and only if, ι↑ (X, ≤) ⊆ τ ⊆ α(X, ≤). Exercise 2.2.5 Show that if a function f : X → Y between topological spaces is continuous, then it is order preserving with respect to the specialization orders on X and Y . Give an example to show that the converse is false. Exercise 2.2.6 Show that if P is a finite partially ordered set, then ι↑ (P) = α(P). Conclude that, in a finite T0 space, any up-set is open. Further show that, for any two partially ordered sets P and Q, a map F : P → Q is order preserving if, and only if, it is continuous with respect to the Alexandrov topologies on P and Q. Note. Using terminology that we will introduce in Definition 5.20, this exercise shows that the category of finite partially ordered sets is isomorphic to the category of finite T0 topological spaces. For further details, see Example 5.21. Exercise 2.2.7 Let X be a compact Hausdorff space and K ⊆ X. Show that K is closed if, and only if, it is compact if, and only if, it is compact saturated.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

42

2 Topology and Order

2.3 Compact Ordered Spaces In this section, we show that there is an isomorphism between certain compact spaces equipped with an order, first introduced by Nachbin, and certain T0 spaces known as stably compact spaces. These spaces provide a well-behaved generalization of compact Hausdorff spaces and also contain the topological spaces dual to distributive lattices, which we call spectral spaces, and which are a main object of study of this book, together with Priestley spaces, their order-topological counterpart. Definition 2.11 An ordered space is a triple (X, τ, ≤) such that • (X, τ) is a topological space, • (X, ≤) is a partially ordered set, and • ≤ ⊆ X × X is closed in the product topology. An ordered space is said to be a compact ordered space provided the underlying topological space is compact. Example 2.12 The set 2 = {0, 1}, equipped with the discrete topology and the usual order, is a compact ordered space. More generally, any finite poset, equipped with the discrete topology, is a compact ordered space. A morphism from an ordered space (X, τX , ≤X ) to an ordered space (Y, τY , ≤Y ) is a function f : X → Y that is both continuous as a map from the space (X, τX ) to (Y, τY ) and order preserving as a map from the poset (X, ≤X ) to (Y, ≤X ). An order homeomorphism between ordered spaces is a morphism that is both a homeomorphism and an order isomorphism; this is sometimes also called isomorphism. For an equivalent definition of order homeomorphism, see Exercise 2.3.2. Proposition 2.13 Let X be an ordered space. Then, the underlying topological space is Hausdorff. Proof Since X is an ordered space, ≤ is closed in X × X equipped with the product topology. Thus, ≥ is also closed in X × X with the product topology and it follows that ΔX = ≤ ∩ ≥ is closed in X × X with the product topology. But this is equivalent to X being Hausdorff (see Exercise 2.1.12).  The following proposition is an important technical tool in the study of compact ordered spaces. Proposition 2.14 Let X be a compact ordered space and C ⊆ X a closed subset of X. Then, ↑C and ↓C are also closed. In particular, ↑x and ↓x are closed for all x ∈ X.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.3 Compact Ordered Spaces

43

Proof If C ⊆ X is closed in X, then C × X is closed in X × X equipped with the product topology. Now, as X is an ordered space, it follows that ≤ is closed and thus (C × X ) ∩ ≤ is closed in X × X. Consider the set π2 [(C × X ) ∩ ≤] = ↑C, where π2 : X × X → X is the projection on the second coordinate. By Proposition 2.5 it follows that it is closed in X. Projecting (C × X ) ∩ ≥ on the second coordinate shows that ↓C is closed. Finally, as X is Hausdorff, it is in particular T1 and thus the singletons x are all closed. It follows that ↑x and ↓x are closed for all x ∈ X.  We can now derive the following very useful order-separation property for compact ordered spaces. Proposition 2.15 Let X be a compact ordered space. For all x, y ∈ X, if x  y, then there are disjoint sets U, V ⊆ X with U an open up-set containing x and V an open down-set containing y. Proof Let x, y ∈ X with x  y. Then, ↑x and ↓y are disjoint. Also, by Proposition 2.14, the sets ↑x and ↓y are closed. Now, since X is compact ordered, it is compact Hausdorff and therefore also normal (see Exercise 2.1.15). Thus, there are open disjoint sets U, V ⊆ X with ↑x ⊆ U and ↓y ⊆ V . Finally, let U  = (↓U c ) c and V  = (↑V c ) c , then one may verify that U  is an open up-set, V  is an open down-set, and we have x ∈ U  and y ∈ V  and U  ∩ V  = ∅.



To any compact ordered space, we now associate two T0 spaces. These spaces are not T1 as long as the order on X is non-trivial. To be specific, if (X, τ, ≤) is a compact ordered space, then we define τ ↑ = τ ∩ U (X, ≤) and τ ↓ = τ ∩ D(X, ≤). In other words, τ ↑ is the intersection of the topology τ and the Alexandrov topology on (X, ≤) and τ ↓ is the intersection of the topology τ and the dual Alexandrov topology on (X, ≤). Accordingly, τ ↑ and τ ↓ are indeed topologies on X. We will often denote the topological space underlying the original ordered space (X, τ, ≤) simply by X, the space (X, τ ↑ ) by X ↑ , and the space (X, τ ↓ ) by X ↓ . Note that, if (X, τ, ≤) is a compact ordered space, then so is (X, τ, ≥). Thus, any property of the spaces X ↑ and their relation to ≤ implies that the order-dual property is true for the spaces X ↓ and we will not always state both. Proposition 2.16 Let (X, τ, ≤) be a compact ordered space. Then, the specialization order of X ↑ is ≤ and in particular X ↑ is a T0 space.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

44

2 Topology and Order

Proof For each x ∈ X, ↓x is closed in (X, τ) and it is a down-set. Therefore, ↓x is closed in X ↑ and it follows that ι↑ (X, ≤) ⊆ τ ↑ . Also, clearly τ ↑ ⊆ U (X, ≤) =  α(X, ≤) and thus the specialization order of X ↑ is ≤ (see Exercise 2.2.4). A crucial fact, given in the following proposition, which will enable us to come back to a compact ordered space X from X ↑ , is the fact that X ↑ and X ↓ are interdefinable by purely topological means without using the data of the original compact ordered space. Proposition 2.17 Let (X, τ, ≤) be a compact ordered space. Then, V ∈ τ↓

⇐⇒

V c ∈ K S(X ↑ ).

Proof Note that C(X ↓ ) = C(X, τ) ∩ U (X, ≤). Also, by definition, K S(X ↑ ) consists of those subsets of X that are both compact with respect to τ ↑ and belong to U (X, ≤). Thus, we need to show that if S ∈ U (X, ≤), then S is closed relative to τ if, and only if, it is compact relative to τ ↑ . Let S ∈ U (X, ≤). If S is closed relative to τ, then S is compact relative to τ (see Exercise 2.1.7). But then it is also compact relative to the smaller topology τ ↑ as required. For the converse, suppose now that S is compact relative to τ ↑ and let y  S. For each x ∈ S, since x  y, by Proposition 2.15, there are disjoint sets Ux, Vx ⊆ X with Ux an open up-set containing x and Vx an open down-set containing y. It follows that {Ux }x ∈S is an open cover of S relative to τ ↑ . Thus, by compactness, there is a finite subset F ⊆ S such that {Ux }x ∈F covers S. Let V  {Vx | x ∈ F}. Then, V is disjoint from the union of {Ux }x ∈F and thus from S. Also, V is open relative to τ ↓ and y ∈ V . That is, we have shown that S is closed relative to τ ↓ .  We are now ready to introduce a class of (unordered) topological spaces called stably compact spaces. Stably compact spaces have a fairly complex definition but, as we will see, they are in fact none other than those spaces which occur as X ↑ for X a compact ordered space. Before we give the definition (Definition 2.18), we need to identify two more properties of spaces, both related to the interaction of compactness and intersection. First, we call a compact space coherent provided the intersection of any two compact-saturated subsets is again compact. A space is called well filtered provided for any filtering collection F of compact-saturated sets and any open set U we have F ⊆U =⇒ there exists K ∈ F such that K ⊆ U. One can show that if X is well filtered, then the collection of compact-saturated subsets of X is closed under filtering intersections (see Exercise 2.3.8). Therefore,

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.3 Compact Ordered Spaces

45

if a space is both coherent and well filtered then the collection of compact-saturated sets is actually closed under arbitrary intersections. Also, notice that compact Hausdorff spaces are both cohereent and well filtered since the compact-saturated sets are just the closed sets. Definition 2.18 A stably compact space is a topological space which is T0 , compact, locally compact, coherent, and well filtered. Example 2.19 Any finite T0 space is stably compact. A particularly important stably compact space is the Sierpinski space S = ({0, 1}, {∅, {1}, {0, 1}}). Note that the topology on S is equal to τ ↑ , where ({0, 1}, τ, ≤) is the compact ordered space described in Example 2.12. Using Proposition 2.17 it is not hard to see that X ↑ is stably compact whenever X is a compact ordered space. We will now show that this is in fact one direction of a one-to-one correspondence between compact ordered spaces and stably compact spaces. To this end we need the notion of the co-compact dual of a topology. Let τ be a topology on a set X. The co-compact dual of τ, denoted τ ∂ , is the topology generated by the complements of compact-saturated subsets of (X, τ). That is, τ ∂ := {K c | K ∈ K S(X, τ)}. The compact-saturated sets are always closed under finite unions, so the collection of their complements is closed under finite intersections and is thus a base for τ ∂ . In the case of a stably compact space, the compact-saturated sets are also closed under arbitrary intersections (see Exercise 2.3.8), so the collection of their complements is already a topology and we have τ ∂ = {K c | K ∈ K S(X, τ)}. Further, we define the patch topology, denoted τ p , obtained from τ to be τ p  τ ∨ τ∂. We can now show how to get back the topology of a compact ordered space X from the topology of X ↑ . Proposition 2.20 Let (X, τ, ≤) be a compact ordered space. Then, (τ ↑ ) ∂ = τ ↓ and (τ ↑ ) p = τ. Proof The first equality is a just a restatement of Proposition 2.17 in terms of the co-compact dual topology. Once we have this, we may observe that Proposition 2.15 tells us, among other things, that τ ↑ ∨ (τ ↑ ) ∂ = τ ↑ ∨ τ ↓ is a Hausdorff topology

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

46

2 Topology and Order

which is contained in τ. But, by Proposition 2.10, if a Hausdorff topology is below  a compact topology, then in fact they are equal, so (τ ↑ ) p = τ as desired. Theorem 2.21 The assignments (X, τ, ≤) → (X, τ ↑ ) and (X, ρ)

→ (X, ρp, ≤ρ )

establish a one-to-one correspondence between compact ordered spaces and stably compact spaces. Proof It is left as Exercise 2.3.11 to show that if (X, τ, ≤) is a compact ordered space, then X ↑ is stably compact. Here, we show that if (X, ρ) is a stably compact space, then (X, ρ p, ≤ρ ) is compact ordered. As a first step, we show that ≤ρ is closed relative to the product topology induced by the patch topology. Let x, y ∈ X with x ρ y. Then, by definition of the specialization order, there is an open set U ∈ ρ with x ∈ U and y  U. By local compactness of ρ there exist V ∈ ρ and K ⊆ X which is compact such that x ∈ V ⊆ K ⊆ U. Since ↑K is also compact and K ⊆ ↑K ⊆ U since U is an up-set in the specialization order, we may assume, without loss of generality, that K ∈ K S(X, ρ). It follows that both V and K c are in ρ p , that x ∈ V , y ∈ K c , and, because V is an up-set and V ∩ K c = ∅, we have (V × K c ) ∩ ≤ρ = ∅. Thus, we have shown that an arbitrary element (x, y) of the complement of ≤ρ lies in a basic open of the product topology which is disjoint from ≤ρ as required. Next, we show that (X, ρp ) is compact. By the Alexander subbase theorem, it suffices to show that covers by subbasic opens have finite subcovers. We use the subbase for ρ p given by S = {K c | K ∈ K S(X, ρ)} ∪ ρ. Now, let C ⊆ S be a cover of X. Define CK S = {K ∈ K S(X, ρ) | K c ∈ C} and Cρ = C ∩ ρ. Then, the fact that C covers X implies that  CK S ⊆ Cρ .

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.3 Compact Ordered Spaces

47

Notice that U = Cρ is open in (X, ρ). Let CK S denote the closure of CK S under finite intersections, that is,   CK S := L | L ⊆ CK S is finite . Note that CK S is, by compactness and coherence, a collection of compact-saturated subsets of (X, ρ), which is moreover filtering. Thus, by well-filteredness, it follows that there is a finite L ⊆ CK S with  L⊆ Cρ .  Now, this means that Cρ is an open cover of the compact-saturated set L. By  compactness there is a finite subcover Cρ of L. This in turn is equivalent to saying that C   {K c | K ∈ L} ∪ Cρ is a finite subcover of the original cover C as required. Thus, it follows that (X, ρp, ≤ρ ) is indeed a compact ordered space whenever (X, ρ) is a stably compact space. Further, the combination of Proposition 2.16 and Proposition 2.20 implies that the composition of the two assignments gives the identity on compact ordered spaces. It remains to show that the reverse composition yields the identity on stably compact spaces. To this end, let (X, ρ) be a stably compact space. Since ρ ⊆ ρp and ρ ⊆ U (X, ≤ρ ), it follows that ρ ⊆ ρ p ∩ U (X, ≤ρ ). The fact that ρ p ∩ U (X, ≤ρ ) ⊆ ρ follows from Exercise 2.3.13.  While the correspondence of the above theorem provides an isomorphism for objects, the natural classes of maps for compact ordered spaces and stably compact spaces are not the same. A natural notion of a structure-preserving map for stably compact spaces is that of a continuous function. For compact ordered spaces the natural notion of a structure-preserving map is that of a function which is simultaneously continuous and order preserving. It is not hard to see that every continuous and order-preserving map between compact ordered spaces is continuous for the corresponding stably compact spaces. However, the converse is not true in general. In fact, the continuous and order-preserving maps between compact ordered spaces correspond to the so-called proper maps between stably compact spaces (see Exercise 2.3.15). We will show in Example 5.22 of Chapter 5 that this correspondence formally yields an isomorphism of categories. We finish this section by recording a useful “translation” between various properties of subsets of a compact ordered space and its corresponding stably compact space. The proof is left as an instructive exercise in applying Theorem 2.21.

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

48

2 Topology and Order

Proposition 2.22 Let (X, τ, ≤) be a compact ordered space with (X, τ ↓ ) and (X, τ ↑ ) the stably compact spaces of τ-open up-sets and τ-open down-sets, respectively. For any subset S of X, (a) (b) (c)

S is saturated in (X, τ ↑ ) if, and only if, S is an up-set in (X, τ, ≤) if, and only if, the complement of S is saturated in (X, τ ↓ ); S is closed in (X, τ ↑ ) if, and only if, S is a closed down-set in (X, τ, ≤) if, and only if, S is compact and saturated in (X, τ ↓ ); and S is compact and open in (X, τ ↑ ) if, and only if, S is a clopen up-set in (X, τ, ≤) if, and only if, the complement of S is compact and open in (X, τ ↓ ).

Exercises for Section 2.3 Exercise 2.3.1 Let (X, τ, ≤) be a triple such that (X, τ) is a topological space and (X, ≤) is a poset. Prove that the following are equivalent: (a) (b)

The order ≤ is closed in (X, τ) × (X, τ). For every x, y ∈ X, if x  y, then there exist open subsets U and V of (X, τ) such that x ∈ U, y ∈ V , and ↑U ∩ ↓V = ∅.

Exercise 2.3.2 Let f : X → Y be a morphism between ordered spaces. Prove that f is an order homeomorphism if, and only if, there exists a morphism g : Y → X such that g ◦ f = idX and f ◦ g = idY . Exercise 2.3.3 Ordered variants of regularity and normality also hold for compact ordered spaces. In particular, prove that if C and D are closed subspaces of a compact ordered space with ↑C ∩ ↓D = ∅, then there are disjoint sets U, V ⊆ X with U an open up-set containing C and V an open down-set containing D. Exercise 2.3.4 Let X be a set and denote by 2 the set {0, 1}. (a) (b)

Show that the set 2X of functions from X to 2 is bijective with P (X ). Show that the product topology on 2X , where 2 is equipped with the Sierpinski topology, translates to the topology on P (X ) given by the subbase {η x | x ∈ X },

(c)

where η x = {S ⊆ X | x ∈ S}, and by the base {η F | F ⊆ X is finite}, where η F = {S ⊆ X | F ⊆ S}. Show that the product topology on 2X , where 2 is equipped with the discrete topology, translates to the topology on P (X ) given by the subbase {η x, μ x | x ∈ X },

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

2.3 Compact Ordered Spaces

49

where μ x = {S ⊆ X | x  S}, and by the base {η F ∩ μG | F, G ⊆ X are finite}, where μG = {S ⊆ X | G ∩ S = ∅}. Note. It will follow from Exercise 2.3.6 that the space 2X , equipped with the pointwise order, is a compact ordered space. Spaces of the form 2X are known as ordered generalized Cantor spaces, and we will encounter them again later in the book, see Exercise 3.3.9 and Definition 4.7. Exercise 2.3.5 Show that R with its usual order and topology is an ordered space. Show that the unit interval [0, 1] is a compact ordered space. Exercise 2.3.6 Show that the set 2 = {0, 1}, equipped with the discrete topology and the usual order, is a compact ordered space. Also, show that compact ordered spaces are closed under an arbitrary Cartesian product, where the product is equipped with the product topology and the coordinate-wise order. Conclude in particular that for any set X, the space 2X is a compact ordered space. Exercise 2.3.7 Let (X, τ, ≤) be a compact ordered space. (a) (b)

Show that the inclusion U (X, ≤) → P (X ) has a lower adjoint given by ↑( ) : P (X )  U (X, ≤) and an upper adjoint given by S → (↓S c ) c . Conclude that Proposition 2.17 implies that the inclusion of K S(X ↑ ) into K S(X ), which is equal to C(X ), has a lower adjoint given by C → ↑C and that the inclusion τ ↑ → τ has an upper adjoint given by U → (↓U c ) c .

Note. We thank Jérémie Marquès for suggesting this reformulation of Proposition 2.17 in terms of upper and lower adjoints on opens and compact-saturated sets. Exercise 2.3.8 (a) Show that if X is a well-filtered space, then the intersection of any filtering collection of compact-saturated sets is again compact saturated. (b) Show that if a collection F of subsets of P (X ) is filtering and closed under  finite intersections (i.e., for any finite S ⊆ F , S ∈ F ), then F is closed under arbitrary intersections. (c) Conclude that if X is well filtered, compact, and coherent, then any intersection of compact-saturated sets is again compact saturated. Exercise 2.3.9 compact.

Show that if X is a compact Hausdorff space, then X is stably

Exercise 2.3.10 2.21.

Deduce Proposition 2.22 from Proposition 2.20 and Theorem

Exercise 2.3.11 compact.

Show that if X is a compact ordered space, then X ↑ is stably

Exercise 2.3.12 Let (X, τ) be a T0 space. Show that ≤τ ∂ = ≥τ .

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

50

2 Topology and Order

Exercise 2.3.13 Let (X, τ) be a stably compact space, B ⊆ τ, and K ⊆ K S(X, τ) such that, for all x, y ∈ X, if x  y, then there exist U ∈ B and K ∈ K such that x ∈ U and y  K. Show that B is a subbase for τ. Exercise 2.3.14 Show that continuous retracts of stably compact spaces are again stably compact. Here a subspace Y of a topological space X is called a (continuous) retract of X provided there exists a continuous function f : X → Y satisfying f (y) = y for all y ∈ Y . Exercise 2.3.15 Let f : X → Y be a continuous function between topological spaces. We call f proper provided the following two properties hold: • ↓ f (C) is closed whenever C ⊆ X is closed; and • f −1 [K] is compact for any K ⊆ Y which is compact saturated. Now, let X and Y be compact ordered spaces and denote the corresponding pair of stably compact spaces by X ↑, X ↓ and Y ↑, Y ↓ , respectively. Further, let f : X → Y be a function between the underlying sets. Show that the following conditions are equivalent: (i) (ii) (iii)

the function f , viewed as a map between compact ordered spaces, is continuous and order preserving; the function f , viewed as a map between the stably compact spaces X ↑ and Y ↑ , is proper; and the function f , viewed as a map between the stably compact spaces X ↓ and Y ↓ , is proper.

Notes for Chapter 2 For readers who need further introduction to topology and for general topology references beyond what we have included here, we recommend a classical book on general topology such as Engelking (1989). For a comprehensive reference on topology and order we recommend Gierz et al. (2003). The correspondence between stably compact spaces and compact ordered spaces in the form given in Theorem 2.21 originates with Gierz et al. (1980). More focused sources presenting the correspondence and the relation with the co-compact dual of a topology are Jung (2004) and Lawson (2011).

https://doi.org/10.1017/9781009349680.003 Published online by Cambridge University Press

3 Priestley Duality

In this chapter, we show how to extend the duality for finite distributive lattices given in Chapter 1 to all distributive lattices. The two key ideas, due to Stone, are to generalize the join/meet-prime elements of the finite case to prime filters/ideals (Section 3.1), and to introduce topology on the structure dual to a distributive lattice (Section 3.2), which leads us to a dual equivalence or duality. The main technical tool is Stone’s prime filter-ideal theorem, Theorem 3.10. In this chapter, we elaborate a modern variant of Stone’s original duality, Priestley duality. The precise connection between Priestley’s duality and Stone’s original duality will be made in Theorem 6.4 in Chapter 6. After treating distributive lattice duality, we show in the final Section 3.3 how the more widely known duality for Boolean algebras follows as an easy consequence. Throughout the chapter, in between introducing the general concepts and proving results about them, we show how to compute dual spaces of several distributive lattices, as “running examples.” We encourage the reader to work through the examples and accompanying exercises in detail, as we believe this is crucial for developing an intuition for dual spaces.

3.1 Prime Filters and Ideals Recall from Section 1.3 that every finite distributive lattice can be represented by the poset of its join-irreducible elements. The following example of an infinite distributive lattice shows why join-irreducibles cannot be used in general. Example 3.1 Consider the lattice L = {⊥} ⊕ (Nop ) 2 , depicted in Figure 3.1. Every non-bottom element (m, n) in this lattice is the join of the two elements (m + 1, n) and (m, n + 1) strictly below it. Therefore, there are no join-irreducibles in L, J (L) = ∅. To find the correct notion that should replace “join-irreducible” in the case of infinite distributive lattices, we need to change our perspective from specific

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

52

3 Priestley Duality (0, 0)

⊥ Figure 3.1 A distributive lattice with no join-irreducible elements.

elements of the lattice to specific subsets of the lattice. If L is a distributive lattice and j ∈ J (L), then j can be uniquely represented by the collection Fj of elements greater than or equal to j. Elements of Fj can be thought of as “approximations” of the element j, which grow in precision as one moves downward in the set Fj ; indeed, j is the infimum of this set Fj . In the case of infinite lattices, while join-irreducibles j themselves may fail to exist (Example 3.1), these “approximating sets” F will still exist. In distributive lattices, where join irreducible and join prime are equivalent notions, there is a simple notion of such an “approximating set,” namely that of a prime filter, defined as follows. Definition 3.2 Let L be a distributive lattice. A subset F of L is called a filter if it is non-empty, an up-set, and for any a, b ∈ L, if a ∈ F and b ∈ F, then a ∧ b ∈ F. A filter is called proper if F  L, or, equivalently, ⊥  F. A filter F is called prime provided that F is proper, and, for any a, b ∈ L, if a ∨ b ∈ F, then a ∈ F or b ∈ F. One way to think of filters in a lattice is that a filter represents an “idealized element,” in much the same way as Noether’s ideals in rings. From this point of view, a non-principal filter stands for a meet that does not exist in the lattice. Note that if S is a subset of a lattice and S ⊆ S , if infima of both sets exist, then    S ≤ S. Thus, since taking a meet over a larger set yields a smaller element, it is natural to postulate the reverse inclusion on filters, as we will do here. The dual notion of ideal in a lattice is introduced in Definition 3.8; there, the order of (non-reversed) subset inclusion is the natural one, since ideals in a lattice stand for an idealized join, and taking joins over larger sets yield larger elements. With the point of view that filters stand for idealized meets, prime filters stand for idealized join-prime elements. More precisely, in finite lattices, prime filters are in one-to-one correspondence with join-prime elements: To any join-prime element j, one may associate the prime filter Fj := ↑j of elements greater than or equal to j (see Exercise 3.1.3). Notice that, also here, j ≤ j  if, and only if, Fj  ⊆ Fj . Thus, this correspondence is an isomorphism of posets if one equips the set of prime

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.1 Prime Filters and Ideals

53

filters with the partial order of reverse inclusion. All this motivates the following definition of a partial order on the set of filters. Definition 3.3 If F and F  are filters in a lattice L, we say that F is below F , F ≤ F  if, and only if, F  is a subset of F. We denote by Filt(L) the poset of filters of L, and by PrFilt(L) the poset of prime filters of L. The following examples of infinite distributive lattices and their posets of prime filters will be used as running examples in this chapter. Example 3.4 Write N ⊕ 1 for the total order on the set N ∪ {ω} which extends N by setting n ≤ ω for all n ∈ N ∪ {ω}. All prime filters of N ⊕ 1 are principal; they are the sets of the form Fn := ↑n for n ∈ (N ⊕ 1) –{0}. Their order is the same as that of N ⊕ 1; this is a highly exceptional case of a distributive lattice that is orderisomorphic to its poset of prime filters, via the isomorphism that sends ω to Fω and n ∈ N to Fn+1 . This lattice and its poset of prime filters are depicted in Figure 3.2. .. .

ω 2 1 0

.. .

Fω F2 F1

Figure 3.2 The distributive lattice N ⊕ 1 and its poset of prime filters.

Now consider the Cartesian product (N ⊕ 1) 2 := (N ⊕ 1) × (N ⊕ 1). Since N ⊕ 1 is a distributive lattice, so is (N ⊕ 1) 2 . The prime filters of (N ⊕ 1) 2 are the principal filters of the form F(n,m) = ↑(n, m), where n = 0 or m = 0 and not both. Indeed, for any (n, m) ∈ (N ⊕ 1) 2 , (n, m) ≤ (n, 0) ∨ (0, m), so if ↑(n, m) is prime, then we must have m = 0 or n = 0. The partial order on the prime filters is the order of the disjoint union, that is, (n, m) ≤ (n , m  ) when either n ≤ n  and m = m  = 0, or n = n  = 0 and m ≤ m . The fact that the prime filters of a Cartesian product split as a disjoint union is not a coincidence (see Exercise 3.1.9). Example 3.5 Consider the totally ordered set N ⊕ Nop , the ordered sum of N and its opposite; that is, the underlying set is {(n, i) | n ∈ N, i ∈ {0, 1}}, and the order is given by (n, i) ≤ (m, j) if, and only if, one of the following holds: • i < j; or • i = j = 0 and n ≤ m; or • i = j = 1 and m ≤ n.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

54

3 Priestley Duality

In this distributive lattice, every (n, i)  (0, 0) is join irreducible, so F(n,i) = ↑(n, i) is a prime filter. Moreover, there is one non-principal prime filter, Fω := {(n, 1) | n ∈ N}. The order ≤ on prime filters is inherited from the order on the lattice, and Fω ≤ F(n,i) if, and only if, i = 1; see Figure 3.3. F(0,1) F(1,1) F(2,1)

(0, 1) (1, 1) (2, 1) .. .

.. . Fω .. .

.. . (2, 0) (1, 0) (0, 0)

F(2,0) F(1,0)

Figure 3.3 The distributive lattice N ⊕ Nop and its poset of prime filters.

The following example shows explicitly that prime filters in distributive lattices generalize prime numbers. They are in fact closely related to prime ideals of rings (see Exercise 3.1.10). Example 3.6 Consider the set N, equipped with the partial order | of divisibility, defined by n | m iff there exists q ∈ N such that m = qn; in particular, n | 0 for all n ∈ N. The partial order (N, |) is a bounded distributive lattice: It has 1 as its bottom, 0 as its top, and for any m, n ∈ N, m ∧ n is the greatest common divisor of m and n, and m ∨ n is the least common multiple of m and n. The prime filters of (N, |) are F0 := {0}, and the sets Fp k := {n ∈ N : pk | n}, for every prime number p and k ≥ 1. It is easy to verify that the Fp k are indeed prime filters (see Exercise 3.1.8); we show that every prime filter is of this form. Suppose that F is a prime filter of (N, |) and F  F0 . Let m be the minimal non-zero element of F with respect to the usual total order on N. We show first that F = ↑m, where the up-set is taken with respect to the divisibility order. Clearly, ↑m ⊆ F. For the converse inclusion, let n ∈ F be arbitrary, n > 0. Then, n ∧ m ∈ F since F is a filter. The greatest common divisor of n and m is non-zero and ≤ m, and therefore equal to m, by minimality of m and thus m | n. We now show that m = pk for some prime p and k ≥ 1. First, note that m > 1 since F is proper. Pick a prime divisor p of m and pick k maximal such that pk | m. Then, m = pk ∨ pmk , so, since F is a

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.1 Prime Filters and Ideals

prime filter, we must have either pk ∈ F or

m pk pk

∈ F. But

55 m pk

< m, so it cannot be

∈ F, and 0 < pk ≤ m, so we get an element of F by minimality of m. Thus, m = pk , again by minimality of m. The partial order on the set of prime filters is given by, for any p, q prime and k,  ≥ 1, Fp k < F0 , and Fp k ≤ Fq  if, and only if, p = q and k ≤ ; see Figure 3.4. F0 .. .

.. . F9

F4

F2

.. . ...

F25

F3

F5

Figure 3.4 The poset of prime filters of (N, |).

Example 3.7 Consider the set L of those subsets of the real unit interval [0, 1] that can be written as a finite union of open rational intervals, that is, as a finite union of sets of the form (p, q), [0, q) or (p, 1], with p, q ∈ Q ∩ [0, 1]. Note that, since a finite intersection of open rational intervals is again an open rational interval, L is a sublattice of P ([0, 1]) under the inclusion ordering, and therefore it is in particular a distributive lattice. There are three kinds of prime filters in the distributive lattice L: • for every q ∈ [0, 1], Fq := {U ∈ L | q ∈ U}; • for every q ∈ [0, 1) ∩ Q, Fq+ := {U ∈ L | for some q  > q, (q, q  ) ⊆ U}; and • for every q ∈ (0, 1] ∩ Q, Fq− := {U ∈ L | for some q  < q, (q , q) ⊆ U}. One may prove (see Exercise 3.1.7) that every prime filter of L is of one of the above forms, that each Fq is maximal in the partial order, while Fq+ and Fq− are strictly below Fq . The partial order is depicted in Figure 3.5. In the rest of this section, we will substantiate the claim that any distributive lattice contains “enough” prime filters. For this purpose, and also for other applications to follow, it will be convenient to introduce the order-dual notion to prime filters: prime ideals. Definition 3.8 Let L be a distributive lattice. A subset I of L is called an ideal if it is non-empty, a down-set, and for any a, b ∈ L, if a ∈ I and b ∈ I, then a ∨ b ∈ I. An ideal I is called proper if I  L, or, equivalently,   I. An ideal I is called

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

56

3 Priestley Duality

F0 F0+

Fq−

q∈Q

r ∈ R–Q

Fq

Fr Fq+

F1 F1−

Figure 3.5 The poset of prime filters of the rational intervals in [0, 1].

prime provided that I is proper, and, for any a, b ∈ L, if a ∧ b ∈ I, then a ∈ I or b ∈ I. Notice that a subset I ⊆ L is a (prime) ideal if, and only if, I is a (prime) filter in L op . Thus, it follows by order-duality from Exercise 3.1.3 that, for a finite lattice L, the prime ideals in L are in a one-to-one correspondence with meet-prime elements of L. This correspondence associates, to any meet-prime element m of L, the prime ideal Im := ↓m of elements less than or equal to m. Notice that m ≤ m  if, and only if, Im ⊆ Im . This motivates the following definition of a partial order on the set of ideals. If I and I  are ideals in a lattice L, we say that I is below I , I ≤ I , if, and only if I is a subset of I . We denote by Idl(L) the poset of ideals of L, and by PrIdl(L) the poset of prime ideals of L. Just as join-prime and meet-prime elements are complementary notions (see Exercise 1.3.7 in Chapter 1), prime filters and prime ideals are complementary, in the following (literal) sense. Lemma 3.9 Let L be a lattice and F ⊆ L. The following are equivalent: (i) (ii) (iii)

The set F is a prime filter. The set I := L – F is a prime ideal. The characteristic function χ F : L → 2, which sends an element a of L to 1 if a ∈ F, and to 0 if a  F, is a lattice homomorphism.

Exercise 3.1.11 asks you to prove Lemma 3.9. We note also that under the equivalence given in this lemma, the partial order on prime filters, which is by definition the reverse subset inclusion order (Definition 3.3), transfers to the usual subset inclusion order on prime ideals. Note also that, for the characteristic functions corresponding to prime filters F and F , we have that F  ≤ F if, and only if, F ⊆ F  if, and only if, χ F ≤ χ F  in the pointwise ordering on functions. The main result about prime filters and prime ideals, which we will prove now, is that there are “enough” of them in any distributive lattice, in the following sense.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.1 Prime Filters and Ideals

57

Theorem 3.10 (Stone’s prime filter-ideal theorem) Let L be a distributive lattice. If F is a filter in L and I is an ideal in L such that F ∩ I = ∅, then there exists a prime filter G in L such that F ⊆ G and G ∩ I = ∅. Proof Consider the partially ordered set S := {G ∈ Filt(L) | F ⊆ G and G ∩ I = ∅}, ordered by inclusion. This set S is non-empty, because it contains F, and if C ⊆ S

is a non-empty chain in S, then G ∈ C G is a filter (see Exercise 3.1.12), and it belongs to S. By Zorn’s lemma (Lemma 2.7), pick a maximal element G in S. We prove that G is prime. By Lemma 3.9, we may show equivalently that J := L – G is a prime ideal. Since G is a filter disjoint from I, we immediately see that J is a down-set containing I, and hence in particular non-empty, and if j1 ∧ j2 ∈ J, then j1 ∈ J or j2 ∈ J. It remains to show that J contains the join of any two of its elements. Let a1, a2 ∈ J. We use the filters generated by G ∪ {a1 } and G ∪ {a2 } (see Exercise 3.1.13). For i = 1, 2, the set Gi := G ∪ {ai }filt = {b ∈ L | there exists g ∈ G such that g ∧ ai ≤ b} is a filter that strictly contains G. By the maximality of G in S, Gi must intersect I non-trivially. Therefore, for i = 1, 2, pick gi ∈ G such that gi ∧ ai ∈ I. Define g := g1 ∧ g2 . Then, using distributivity, g ∧ (a1 ∨ a2 ) = (g ∧ a1 ) ∨ (g ∧ a2 ) ≤ (g1 ∧ a1 ) ∨ (g2 ∧ a2 ). Since gi ∧ ai ∈ I for i = 1, 2, it follows from this that g ∧ (a1 ∨ a2 ) ∈ I. Therefore, since J contains I, we have g ∧ (a1 ∨ a2 ) ∈ J. Since G is a filter, we have g ∈ J or a1 ∨ a2 ∈ J. However, g ∈ J is impossible since g1 ∈ G and g2 ∈ G. Thus,  a1 ∨ a2 ∈ J, as required. We make note of the fact that the proof of Theorem 3.10 relies on Zorn’s lemma, which we already encountered in the context of the Alexander subbase theorem in Chapter 2. In fact, in Zermelo–Fraenkel set theory without choice, the statement of Theorem 3.10 is strictly weaker than the axiom of choice. It is equivalent to both the ultrafilter theorem for Boolean algebras and to the Alexander subbase theorem (see, for example, Howard and Rubin, 1998). From Theorem 3.10, we obtain the following representation theorem for distributive lattices, due to Stone (1937b). Theorem 3.11 (Stone representation for distributive lattices) Let L be a lattice. The function  : L → D(PrFilt(L)) (−) a →  a := {F ∈ PrFilt(L) | a ∈ F}

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

58

3 Priestley Duality

 is injective if, and only if, is a well-defined lattice homomorphism. Moreover (−) L is distributive. In particular, any distributive lattice L embeds into the lattice of down-sets of the poset PrFilt(L).  is a well-defined lattice Proof We leave it as Exercise 3.1.14 to prove that (−) homomorphism. If it is moreover injective, then L is distributive, as L is then isomorphic to a sublattice of a distributive lattice. Conversely, suppose that L is distributive and let a, b ∈ L be such that a  b. Then, the filter F := ↑a generated by a is disjoint from the ideal I := ↓b generated by b. By Theorem 3.10, pick a prime filter G containing F and disjoint from I. Then, G ∈  a – b, so  a b.   here as the notation we used in Section Note that we use the same notation, (−), 1.3 for the function from a finite lattice L to the lattice of down-sets of joinirreducible elements of L. Indeed, if L is a finite lattice, then for any a ∈ L, the set  a defined here in Theorem 3.11 and the set  a in Section 1.3 correspond to each other, under the correspondence between prime filters and join-prime elements that holds in finite lattices (see Exercise 3.1.3). Theorem 3.11 is less satisfactory than the representation theorem for finite distributive lattices, because it does not give an isomorphism, but only a lattice embedding of a distributive lattice into a lattice of down-sets. In order to get an isomorphism, and then a full duality, we will introduce a topology on the set of prime filters in the next section.

Exercises for Section 3.1 Exercise 3.1.1 equivalent: (a) (b) (c)

Let L be a lattice and F ⊆ L. Prove that the following are

The set F is a filter.  For every finite S ⊆ L, we have S ∈ F if, and only if, S ⊆ F. The characteristic function χ F : L → 2 of F preserves finite meets.

Exercise 3.1.2 Let L be a lattice. (a) (b)

Prove that a subset I of L is an ideal if, and only if, I is a directed down-set. Conclude that F ⊆ L is a filter if, and only if, F is a filtering up-set.

Exercise 3.1.3 Let L be a lattice. (a) (b) (c)

Prove that, for any a ∈ L, the set Fa := ↑a = {b ∈ L | a ≤ b} is a filter. Prove that, for any join-prime element j of L, the set Fj = ↑j is a prime filter. Prove that, for any prime filter F, if the infimum of F exists in L and belongs  to F, then jF := F is join prime.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.1 Prime Filters and Ideals

(d)

59

Prove that, if L is finite, then the assignments j → Fj and F → j F constitute a well-defined bijection between the poset of join-prime elements of L and the poset of prime filters of L, ordered by reverse inclusion.

Exercise 3.1.4 Prove that every prime filter of the lattice N ⊕ Nop of Example 3.5 is either principal or equal to Fω . Exercise 3.1.5 Prove that every prime filter of the lattice (N ⊕ 1) 2 of Example 3.4 is principal, and thus, by the argument given in that example, of the stated form. Exercise 3.1.6 Consider the total order Q ∩ [0, 1]. Show that there are three types of prime filters in this distributive lattice: • for every q ∈ Q ∩ (0, 1], the principal prime filter Fq = ↑q; • for every q ∈ Q ∩ [0, 1), the prime filter G q := ↑q –{q}; and • for every irrational r ∈ (0, 1), the prime filter Hr := {q ∈ Q ∩ [0, 1] | r < q}. Make a diagram of the partial order on these prime filters. Exercise 3.1.7 Show that the poset of prime filters of the lattice of finite unions of rational intervals from Example 3.7 is as stated. You may proceed as follows: (a) (b)

(c) (d) (e)

Show that, for every q ∈ [0, 1], Fq is a prime filter which is minimal in the inclusion ordering (and thus maximal in the poset of prime filters). Show that, for every prime filter F, there exists a unique q ∈ [0, 1] such that Fq ⊆ F. Hint. For existence, consider the intersection of the closures of the U ∈ F and use the compactness of [0, 1] in the topology generated by L. For uniqueness, use that any two distinct points in [0, 1] have a rational point in between. Show that, for every irrational r ∈ (0, 1), if F is a prime filter and Fr ⊆ F, then F = Fr . Show that, for every rational q ∈ (0, 1), Fq− and Fq+ are maximal prime filters in the inclusion order. Using the previous item, show that, if F is a prime filter and Fq  F for some q ∈ (0, 1), then F = Fq− or F = Fq+ .

Exercise 3.1.8 Prove that, for any prime number p and k ≥ 1, the set {n ∈ N : pk | n} is a prime filter in (N, |). Exercise 3.1.9 Let L and M be distributive lattices. Prove that the prime filters of L × M are exactly the filters that are either of the form F × M for some prime filter F of L, or of the form L × G for some prime filter G of M. Conclude that the poset of prime filters of L × M is the disjoint union of the posets of prime filters of L and of M.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

60

3 Priestley Duality

Exercise 3.1.10 (This exercise requires familiarity with basic ring theory; see the article Banaschewski (1996) and the monograph Dickmann et al. (2019) for more about the link between commutative algebra and lattice theory.) Let R be a commutative ring with unit. Recall that a ring ideal I of R is a subgroup of R such that, for every i ∈ I and r ∈ R, ri is in I. For any subset S ⊆ R, there is a smallest ring ideal containing S, denoted here S R , and consisting of all elements n r i si for some r 1, . . . , r n ∈ R and s1, . . . , s n ∈ S. A ring that can be written as i=1 ideal I is called prime if I  R and, whenever r, s ∈ R such that r s ∈ I, either r ∈ I or s ∈ I. A ring ideal I is called radical if, for every r ∈ R, if r n ∈ I for some n ≥ 1, then r ∈ I. (In this exercise, we use the expressions “(prime) ring ideal” and “(prime) lattice ideal” to avoid confusion between the notions. By “radical ideal” we always mean radical ring ideal.) (a)

(b)

(c)

Prove that, for any set S, the set √ S := {r ∈ R | r n ∈ SR for some n ≥ 1} is the smallest radical ideal containing S, called the radical ideal generated by S. √ By a finitely generated radical ideal, we mean an ideal of the form S, where S is a finite subset of R. Prove that the collection RIdl f g (R) of finitely generated radical ideals of R is a distributive lattice under the inclusion order. Prove that, if P is a prime ring ideal of R, then ϕ(P) := {I ∈ RIdl f g (R) | I ⊆ P}

(d)

is a prime lattice ideal of RIdl f g (R). Prove that ϕ is an order-isomorphism between the prime ring ideals of R and the prime lattice ideals of RIdl f g (R).

Exercise 3.1.11 Prove Lemma 3.9. Exercise 3.1.12 Let D ⊆ Filt(L) be a directed collection of filters in the inclusion order, that is, D is non-empty, and if G1, G2 ∈ D, then there exists G0 ∈ D such

that G1 ∪ G2 ⊆ G0 . Prove that G ∈ D G is a filter. Conclude in particular that the union of a non-empty chain of filters is a filter. Exercise 3.1.13 Let L be a lattice. (a)

(b)

Prove that any intersection of a collection of filters is a filter. Conclude that, in particular, for any S ⊆ L, there exists a smallest filter containing S. We refer to this filter as the filter generated by S and denote it by Sfilt . Prove that, for any S ⊆ L,    T ≤b . Sfilt = b ∈ L | there exists finite T ⊆ S such that

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.2 Priestley Duality

(c)

61

Prove that, if G is a filter and a ∈ L, then G ∪ {a}filt = {b ∈ L | there exists g ∈ G such that g ∧ a ≤ b}.

(d)

Conclude that, if I is a down-set which intersects G ∪ {a}filt non-trivially, then there exists g ∈ G such that g ∧ a ∈ I. Formulate and prove analogous results for the ideal generated by a set S, notation Sidl .

 defined in Theorem Exercise 3.1.14 Prove that, for any lattice L, the function (−) 3.11 is a well-defined lattice homomorphism. Exercise 3.1.15 Let L := {a ⊆ N | a is finite or a = N}, the distributive lattice of finite subsets of N and N itself. (a) (b) (c) (d)

Show that, for each n ∈ N, the set x n := {a ∈ L | n ∈ a} is a prime filter of L. Show that x ∞ := {N} is a prime filter of L. Show that PrFilt(L) = {x n | n ∈ N} ∪ {x ∞ }. Explicitly describe the partial order ≤ on PrFilt(L).

Exercise 3.1.16 A subset a of a set X is called co-finite (in X) if X – a is finite. Let M := {a ⊆ N | a is finite or co-finite}, the distributive lattice of finite or co-finite subsets of N, and let L be the lattice defined in Exercise 3.1.15. (a) (b) (c)

Prove that, for any lattice homomorphism h : L → 2, there exists a unique lattice homomorphism h  : M → 2 that extends h. Hint. Use Exercise 1.2.11(b). Write down an explicit bijection between PrFilt(L) and PrFilt(M). Hint. Use item (a) and Lemma 3.9. Describe the partial order ≤ on PrFilt(M). Which of the two directions of the bijection in the previous item is order preserving?

Exercise 3.1.17 Deduce from Stone’s prime filter-ideal theorem that any filter F in a distributive lattice L equals the intersection of all prime filters G that contain F.

3.2 Priestley Duality In this section, we introduce a topology on the poset PrFilt(L) that makes it into a special kind of compact ordered space, called a Priestley space, after H. A. Priestley, who introduced this topology in Priestley (1970). We will then show that the poset PrFilt(L), equipped with this topology, completely represents the distributive lattice L. Finally, we will also prove that homomorphisms between distributive lattices correspond to continuous order-preserving functions between their dual spaces.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

62

3 Priestley Duality

The Priestley Dual Space of a Distributive Lattice Definition 3.12 Let L be a distributive lattice. The Priestley topology, τ p , on the set PrFilt(L) is the topology generated by the subbase S := { a | a ∈ L} ∪ {( b) c | b ∈ L}. This definition makes PrFilt(L) into a special kind of compact ordered space, as we prove now. Proposition 3.13 Let L be a distributive lattice. (a) (b)

The topology τ p on PrFilt(L) is compact. For any F, G ∈ PrFilt(L), if F  G, then there exists a clopen down-set U in PrFilt(L) such that G ∈ U and F  U.

In particular, (PrFilt(L), τ p, ≤) is a compact ordered space. Proof (a) By the Alexander subbase theorem (Theorem 2.6), it suffices to prove that any cover C ⊆ S of PrFilt(L) by open sets from the subbase has a finite subcover. Let S, T ⊆ L be such that    ( b) c . a∪ PrFilt(L) = a ∈S

b ∈T

Rephrasing slightly, this equality says that every prime filter G of L that contains the set T must intersect the set S non-trivially. In particular, writing F for the filter generated by T and I for the ideal generated by S, every prime filter G of L that contains F must intersect I non-trivially. By Theorem 3.10, the intersection of F and I is non-empty, so pick c ∈ F ∩ I. By Exercise 3.1.13, pick finite T  ⊆ T and finite     S  ⊆ S such that T  ≤ c and c ≤ S . In particular, T  ≤ S , and hence,

 is a lattice homomorphism (Theorem 3.11), we have b ∈T   b ⊆ a ∈S  since (−) a. Rewriting this subset inclusion, we conclude that    ( b) c . a∪ PrFilt(L) = a ∈S

b ∈T 

(b) Let F, G ∈ PrFilt(L) be such that F  G. By definition, there exists a ∈ L such that a ∈ G and a  F. Thus, U :=  a is a clopen down-set that contains G but not F. By item (b) and Exercise 2.3.1, (PrFilt(L), ≤, τ p ) is in particular a compact ordered space.  Proposition 3.13 gives rise to the following definitions. Definition 3.14 An ordered topological space (X, τ, ≤) is called totally orderdisconnected (TOD) if, for every x, y ∈ X, whenever x  y, there exists a clopen

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.2 Priestley Duality

63

down-set U in X such that y ∈ U and x  U. A compact ordered space which is totally order-disconnected is called a Priestley space, sometimes also a CTOD space. The TOD property yields the following stronger version of the order normality property of compact ordered spaces (Exercise 2.3.3) for Priestley spaces. Proposition 3.15 Let C and D be closed sets in a Priestley space X. If ↓C∩↑D = ∅, then there exists a clopen down-set K ⊆ X such that C ⊆ K and K ∩ D = ∅. A proof of Proposition 3.5 is outlined in Exercise 3.2.7. We now examine some Priestley topologies in our running examples. Example 3.16 Recall from Example 3.4 the lattice N ⊕ 1 and its poset of prime n is the finite subset filters X := {F1, F2, . . . } ∪ {Fω }, see Figure 3.2. For any n ∈ N,   = X. It follows that, for every 1 ≤ n < ω, the singleton {Fn } {Fi | 1 ≤ i ≤ n}, and ω is equal to  n – n − 1, and is therefore clopen in the Priestley topology. Therefore, every subset of X –{Fω } is open. If, on the other hand S ⊆ X and Fω ∈ S, then S is open if, and only if, S is co-finite (see Exercise 3.1.16). This topology is homeomorphic to the one-point compactification of the discrete countable space {F1, F2, . . . }, also see Exercise 3.2.2 and Example 3.46. As shown in Example 3.4, the lattice (N ⊕ 1) 2 has as its dual poset the disjoint (n, m) is equal to union of X with itself. For (n, m) ∈ (N ⊕ 1) 2 , the clopen down-set  {F(n,0) | n  ≤ n} ∪ {F(0,m ) | m  ≤ m}. It follows that the topology is also the disjoint union of the topologies on the two copies of X, that is, a subset U ⊆ X + X is open if, and only if, its intersection with both the “left” and “right” copy of X is open (see Exercise 3.2.3). Example 3.17 We continue Example 3.5 of the lattice N ⊕ Nop , whose dual poset was depicted in Figure 3.3. For n ∈ N, we have that  (n, 0) = {F(k,0) | k ≤ n} and  (n, 1) = {F(k,1) | k ≥ n} ∪ {Fω } ∪ {F(m,0) | m ∈ N}. It follows that, for n ≥ 1,  {F(n,0) } = (n, 0) – (n + 1, 0) and, for n ≥ 0,  1) – (n − 1, 1), {F(n,1) } = (n, so that again all the singleton sets except for {Fω } are open in the Priestley topology. As in the previous example, one may again prove that the topology is homeomorphic

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

64

3 Priestley Duality

to the one-point compactification of the discrete countable space {F(k,i) | k ∈ N, i ∈ {0, 1}}. The only thing distinguishing this space from the space X in the previous example is the partial order. Example 3.18 We continue Example 3.6 of the lattice of divisibility (N, |), whose dual poset was depicted in Figure 3.4. Clearly,  0 is the entire space, and for any n ≥ 1, we have  n = {Fp k : pk | n, p prime, k ≥ 1}. In particular, for any prime p and k ≥ 1, we have that k−1, pk – p {Fp k } = 

n is finite for every n ≥ 1, it follows so all singletons except {F0 } are clopen. Since  again that a clopen set K in the Priestley topology contains F0 if, and only if, it is co-finite, and the topology on the dual poset is again the one-point-compactification of a discrete countable space. Example 3.19 We continue Example 3.7 of the lattice of finite unions of rational intervals, whose dual poset was depicted in Figure 3.5. As a useful notation for the rest of this example (also see Notation 3.20), for every r ∈ [0, 1], we denote the prime filter Fr defined in Example 3.7 simply by r, and we write r + and r − for the prime filters that were denoted by Fr+ and Fr− , respectively, whenever they exist. Also, for every r ∈ [0, 1], let us write Xr for the set of prime filters “located” at r, that is, ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ Xr := ⎨ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩

{r } {r, r +, r − } {0, 0+ } {1, 1− }

if r if r if r if r

is irrational, ∈ (0, 1) ∩ Q, = 0, = 1.

Now, toward describing the Priestley topology on the poset X = that, for any q ∈ Q ∩ [0, 1],   Xr ∪ {q− }, [0, q) = r n. Conclude that x has a cover y that is not in U, thus showing that U is not an up-set in (X, ). Suppose that F = ∅ and show that U is an up-set of (X, ) if, and only if, G is an initial segment of N.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

3.3 Boolean Envelopes and Boolean Duality

(3)

(4)

85

Show that a clopen subset of X is a down-set with respect to  if, and only if, it is a down-set with respect to ≤ and thus conclude that (X, ) is not a Priestley space. Find points x and y in X showing that (X, ) does not have the TOD property.

Notes for Chapter 3 The influential American mathematician Marshall H. Stone introduced Stone duality for both Boolean algebras and bounded distributive lattices in a series of papers published in the 1930s (Stone, 1934, 1935, 1936, 1937a, 1937b, 1938). Priestley duality (Priestley, 1970) recasts Stone duality for bounded distributive lattices in terms of Nachbin’s theory of ordered spaces (Nachbin, 1964), and identifies an isomorphic dual category to the one given by Stone (1937b). See also results in this direction by Nerode (1959). The connection between Stone’s original duality and Priestley’s variant has been explained in several places in the literature, for example, Cornish (1975) and Fleisher (2000). While Stone’s representation theorem and duality for Boolean algebras (Stone, 1936) are widely known, his results for distributive lattices, which include both a representation theorem and a duality, have recently received less attention in the literature. One reason for this is the fact that the spaces in Stone’s duality for distributive lattices (Stone, 1937b) are non-Hausdorff. Priestley’s work used a much nicer class of spaces, at the expense of equipping the space with a partial order. In this book, we decided to introduce Priestley’s duality first, as it nicely exhibits the interplay between order and topology that is central to the field. We make the connection with Stone’s original duality in Chapter 6, see Theorem 6.4. While the main results of this chapter are well known, the notation of a “neutral Priestley dual space” is more recent and originates with the papers Gehrke and Priestley (2006, 2007). This notation is particularly useful for studying duality for additional operations which are order preserving in some coordinates and order reversing in others, as we will begin to do in Chapter 4, as well as for studying operations that preserve or reverse both join and meet in each coordinate, which was the subject of the aforementioned papers.

https://doi.org/10.1017/9781009349680.004 Published online by Cambridge University Press

4 Duality Methods

In this chapter, we build on the Priestley duality of Chapter 3 to develop methods for analyzing distributive lattices and the morphisms between them, with applications to logical systems associated to them. We first show in Section 4.1 how Priestley duality allows us to easily compute free distributive lattices and Boolean algebras, corresponding to normal forms for propositional logic formulas. In Section 4.2, we establish correspondences between sublattices of a lattice and quotients of its dual space, and also between quotients of a lattice and closed subspaces of its dual space. These Galois-type correspondences are often instrumental for computing dual spaces of lattices in concrete cases; we will give a few basic examples already in this chapter and we will make full use of these correspondences in Chapters 7 and 8. Sections 4.3 and 4.5 show how duality theory can treat classes of functions that are more general than the lattice homomorphisms that appear in Priestley duality: operators. We first treat unary operators (i.e., functions preserving only finite meets or finite joins, but not both) in Section 4.3, and then in Section 4.5 we focus on a particularly relevant case of operators of arity 2, namely operators of implication type. This case is sufficiently complex to show the flavor of the fully general case, while avoiding notational difficulties when dealing with operators of general arity. In Sections 4.4 and 4.6, we give two classical applications of operator duality, namely to deduce Kripke models for modal logic, and to obtain Esakia duality for Heyting algebras, the algebraic structures for intuitionistic propositional logic. Duality for operators will also be applied in both Chapters 7 and 8; in particular, Section 4.5 contains some forward pointers to where this theory is relevant in both of those chapters. The final section, Section 4.7, contains an application of discrete duality which characterizes the Boolean subalgebra generated by a sublattice. This result, which is of a more combinatorial flavor, will be used in Chapter 8. The methods we develop in this chapter can also be viewed in a more abstract categorical form, and we will do so for some of them in the next chapter,

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.1 Free Distributive Lattices

87

Chapter 5. It is useful to first see them “in action:” this chapter shows the use of individual instruments, while the next chapter shows how they all fit together in an ensemble.

4.1 Free Distributive Lattices In this section, we apply Priestley duality to give concrete descriptions of free distributive lattices over sets. We also deduce a description of the free Boolean algebra over a set as a corollary. We will point out some connections to propositional logic. While our discussion here focuses on free distributive lattices, a large part of the development in this section is an instance of a much more general universal algebra, developed by Birkhoff; a standard reference for this is the textbook by Burris and Sankappanavar (2000). This section does not require knowledge of universal algebra, but we provide some pointers for the interested reader. In Example 5.17, we will relate the definition of free objects to the categorical notion of adjunction, to be introduced in Chapter 5. The construction of free algebras is an important tool in universal algebra. The reason for this is Birkhoff’s variety theorem, which links axiomatization by equations with the model-theoretic constructions of quotients, subalgebras, and Cartesian products (see Burris and Sankappanavar 2000, Theorem 11.9). The technical crux of this theorem is the fact that if a class of algebras is axiomatized by equations, then it contains free algebras over any set (Burris and Sankappanavar, 2000, Theorem 10.12), and any other algebra of the class is a quotient of a free algebra (Burris and Sankappanavar, 2000, Corollary 10.11). While this is interesting, free algebras are also often notoriously difficult to understand and transferring information to quotient algebras may also be challenging. We will see here that, in the specific case of distributive lattices, free distributive lattices are quite simple to understand dually, as their dual spaces are Cartesian products of the two-element Priestley space. Combined with the methods of Section 4.2, which will allow us to describe quotients of distributive lattices dually as closed subspaces, the free algebras from Birkhoff’s variety theorem provide a powerful tool for distributive-lattice-based algebras. We will see this in action, for example, in the applications in Section 7.4. We now define what it means for a distributive lattice to be free. This definition will look familiar if you have previously encountered, for example, free groups or free semigroups. It is also similar to the definition of Boolean envelope, Definition 1.19. The proper general setting for these definitions will be developed in Section 5.2 in the subsection on adjunctions, see Example 5.17(a) in particular. Definition 4.1 Let V be a set, F a distributive lattice, and e : V → F a function. Then, F is said to be free over V via e provided that, for every distributive lattice

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

88

4 Duality Methods

L and every function f : V → L, there exists a unique lattice homomorphism f¯ : F → L such that f¯ ◦ e = f , that is, such that the following diagram commutes: F e

f¯

L

f

V. The property expressed by the diagram is called a universal property of (F, e) with respect to distributive lattices, and a function e : V → F as in this definition is called a universal arrow. Given a universal arrow e : V → F and f : V → L with L a distributive lattice, we call the homomorphism f¯ : F → L the unique extension of f along e. Toward the general construction of free distributive lattices, we make two basic observations about its definition, which in particular justify speaking of “the” free distributive lattice over a set V . Proposition 4.2 Let V be a set. A free distributive lattice over V is unique up to isomorphism. That is, if e : V → F and e  : V → F  are two universal arrows, then there exists a unique lattice isomorphism ϕ : F → F  such that ϕ ◦ e = e . Proof Suppose e : V → F and e  : V → F  are universal arrows. Denote by ϕ : F → F  the unique extension of e  along e, and by ψ : F  → F the unique extension of e along e . Then, the composite function ψ ◦ ϕ : F → F is an extension of e along e, that is, it is a homomorphism with the property that ψ ◦ ϕ ◦ e = e, since, by definition, ϕ ◦ e = e  and ψ ◦ e  = e. But idF : F → F is also an extension of e along e, so by uniqueness, we have ψ ◦ ϕ = idF . By symmetry, ϕ ◦ ψ = idF  .  Thus, ϕ is a lattice isomorphism, and ϕ ◦ e = e  by construction, as required. Notation For V a set, we denote by FDL (V ) the free distributive lattice over V , and by e : V → FDL (V ) the accompanying universal arrow, if they exist. By Proposition 4.2, this notation is well defined, if we consider isomorphic lattices as the same. Proposition 4.3 Suppose FDL (V ) is the free distributive lattice over V , and e : V → FDL (V ) the accompanying universal arrow. Then, FDL (V ) is generated by e[V ]. Proof Denote by L the sublattice of FDL (V ) generated by e[V ]. Denote by p : FDL (V ) → L the unique extension along e of the function e  : V → L, defined as the co-restriction to L of the function e. Write i : L → FDL (V ) for the

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.1 Free Distributive Lattices

89

inclusion homomorphism; the various morphisms are depicted in the following diagram. i FDL (V )

L e

p e

V. Note, similarly to the proof of Proposition 4.2, that i ◦ p is an extension of e along e, and therefore i ◦ p = idFDL (V ) . Hence, i is surjective, proving that L = FDL (V ).  Note that we have not yet established that a free distributive lattice and accompanying universal arrow actually exist for every set V . It is possible to give very general arguments for the existence, because distributive lattices have a finitary axiomatization, see, for example, Burris and Sankappanavar (2000, Section II.10). Here we give two different concrete constructions of the free distributive lattice over V , one coming from the general algebraic considerations, the other using Priestley duality, and thus specific to distributive lattices. The advantage of the second construction is that it gives a concrete representation of the free distributive lattice as clopen down-sets of a particular Priestley space. This representation is often useful in applications; we will in particular make use of it in Chapters 6 and 7. For the algebraic construction, let V be a set. Since the free distributive lattice is generated by the image under the universal arrow, it makes sense that we can build it by making all possible well-formed expressions over V and then take a quotient, as we will do now. A lattice term with variables in V is a well-formed expression built from V using the operation symbols ∨, ∧, , and ⊥. For example, a, (a ∨ b) ∧ a, and ⊥ ∨ a are examples of lattice terms with variables in {a, b}; while these three terms are distinct syntactic objects, they clearly should be considered “equivalent” from the perspective of (distributive) lattices. In fact, the universal property tells us that the free distributive lattice, if it exists, should have every V -generated distributive lattice as a quotient. We will now define an equivalence relation ≡ on the set of terms which makes this idea precise. We write T (V ) for the set of all lattice terms with variables in V . Note that, given a function f : V → L with L any structure equipped with operations ∨, ∧, , and ⊥, we may inductively define an interpretation function f˜ : T (V ) → L, by f˜(v) := f (v) for v ∈ V , f˜() := , f˜(⊥) := ⊥, f˜(t ∨ t  ) := f˜(t) ∨ f˜(t  ), and f˜(t ∧ t  ) := f˜(t) ∧ f˜(t  ) for any terms t, t . Now consider the equivalence relation ≡ on T (V ), defined by t ≡ t  if, and only if, for every function f : V → L, with L a distributive lattice, we have f˜(t) = f˜(t  ). The crucial insight is now that the set T (V )/≡ of

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

90

4 Duality Methods

equivalence classes of lattice terms naturally admits the structure of a distributive lattice, as follows. First, note that the equivalence relation ≡ is congruential on T (V ), that is, if a ≡ a  and b ≡ b, then a∧b ≡ a  ∧b and a∨b ≡ a  ∨b, as is easily verified using the fact that f˜ in the definition of ≡ preserves the operations ∨ and ∧ (see Exercise 4.1.4). Therefore, we have well-defined operations on T (V )/≡ given by  := []≡, ⊥ := [⊥]≡, [a]≡ ∨ [b]≡ := [a ∨ b]≡, and [a]≡ ∧ [b]≡ := [a ∧ b]≡, for any a, b ∈ T (V ). To see that (T (V )/≡, , ⊥, ∨, ∧) is a distributive lattice under these operations, one may verify that the defining equations from Section 1.2 hold in T (V )/≡ (see Exercise 4.1.4). Finally, define e : V → T (V )/≡ by e(v) = [v]≡ . Proposition 4.4 The distributive lattice T (V )/≡ is free over V via the universal arrow e. Proof Let f : V → L be a function to a distributive lattice L. The interpretation function f˜ : T (V ) → L has the property that f˜ ◦ e = f , and, for any t ≡ t , we have f˜(t) = f˜(t  ), by definition of ≡. Therefore, the function f¯ : T (V ) → L, defined by f¯([t]≡ ) := f˜(t) is well defined. Observe that the function f¯ is a homomorphism by definition of the operations on T (V )/≡, and it is an extension of V . We leave uniqueness of f¯ as an exercise (see Exercise 4.1.4).  Remark 4.5 The above construction is essentially that of the Lindenbaum–Tarski algebra for a propositional logic without negation over variables V . A completely analogous development, replacing DL by BA throughout, gives a construction of the free Boolean algebra over a set of variables; also see Corollary 4.12. Before taking on the general duality-theoretic construction, let’s look at a small example. Example 4.6 Let V = {p, q}. We get lattice terms ⊥, , p, q, p ∧ q, and p ∨ q, and it is not too difficult to see that, in any distributive lattice, the interpretation of all other terms must be equal to one of these. That is, FDL ({p, q}) must be the lattice depicted on the left in Figure 4.1. There, the four join-primes of FDL ({p, q}) are identified as the circled nodes. In the figure on the right, we depict the dual poset as (2, ) {p,q } , that is, the set of functions from {p, q} to 2, ordered pointwise, where the order on the set 2 is given by 1  0. Each function f : {p, q} → 2 is denoted as a pair f (p) f (q), for example, 10 is the function sending p to 1 and q to 0. The idea behind this representation of the dual poset of FDL ({p, q}) is the following. As we know from Definition 3.21, we may choose to see the dual poset of FDL ({p, q}) as HomDL (FDL ({p, q}), 2) in the reverse pointwise order. Since FDL ({p, q}) is generated by p and q, each homomorphism h : FDL ({p, q}) → 2 is

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.1 Free Distributive Lattices

91

 00

p∨q q

p

10

01

p∧q 11 ⊥ Figure 4.1 The free distributive lattice on two generators, and its dual poset.

totally determined by its restriction to {p, q}. Also, by the universal property, every function from {p, q} to 2 is the restriction of such a homomorphism. For example, the element p in the lattice corresponds to the down-set {10, 11} of the dual poset while the element ⊥ corresponds to the empty down-set, and the element p∨q corresponds to the down-set {10, 01, 11}. Thus, we see that the concrete incarnation of FDL ({p, q}) obtained via duality is given by the universal arrow   e : {p, q} → D (2, ) {p,q } , p → ↓ χ p, where χ p is the characteristic function of the singleton {p} and the order on 2 is given by 1  0. From a logic perspective, the elements of the dual poset correspond to lines in a truth table for propositional logic on two variables, and the subset associated to a formula ϕ is the set of lines in the truth table at which ϕ is true. This is precisely as conceived by logicians going back to the fundamental work by Boole (1847). We here obtained the universal arrow e as a map to the down-sets of 2 {p,q } in the pointwise order relative to the “upside-down” order on 2, in which 1  0. Note that this map may alternatively be described using up-sets and the poset (2, ≤), where ≤ is the usual order in which 0 ≤ 1, via the map   {p, q} → U 2 {p,q } , p → ↑ χ p . These are two alternate descriptions of one and the same concrete incarnation of the free distributive lattice over the set {p, q}. Our reason for sticking with the “upside-down” version of 2 is that it here plays the role of the poset dual to the free distributive lattice on a single generator, which is a three-element chain, see Remark 4.10 and Exercise 5.3.13. We now proceed to make the corresponding argument for an arbitrary set V . For a set V , let 2V denote the set of functions from V to 2, which may be viewed as

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

92

4 Duality Methods

the |V |-fold product of the two-element set 2. Recall that we already encountered topologies on the set 2V in Exercises 2.3.4 and 3.3.9. We will show in Proposition 4.8 that the dual space of FDL (V ) is order-homeomorphic to the following ordered generalized Cantor space. Definition 4.7 Let V be a set. We define the topology π on 2V to be the product topology, where 2 carries the discrete topology. We define the partial order  on 2V , for x, y ∈ 2V , by f  g ⇐⇒ for all v ∈ V, if g(v) = 1, then f (v) = 1. The ordered topological space (2V , π, ) is called the ordered generalized Cantor space over V . The topological space (2V , π) is called generalized Cantor space, and (2N, π) is the (classical) Cantor space. Note that, by definition of the product topology, the topology π on 2V has as a subbase the sets of the form v → a := { f ∈ 2V | f (v) = a}, where v ranges over the elements of V and a = 0, 1. Note also that  is the pointwise order on 2V with respect to the order on 2 in which 1  0, as in Example 4.6; see Remark 4.9 for an alternative description using up-sets and the order on 2 in which 0 ≤ 1. Proposition 4.8 Let V be a set. The Priestley dual space of the free distributive lattice over V is order-homeomorphic to the ordered generalized Cantor space over V , and the function s : V → ClpD(2V ), which sends v ∈ V to the clopen down-set s(v) := v → 1, is a universal arrow. Our proof here uses the algebraic fact, already proved above, that there exists a free distributive lattice over V . It is possible to give a proof “from scratch” that does not use this fact (see Exercise 4.1.6). Proof Let e : V → F be a free distributive lattice over V and denote by (X, τ, ≤) the Priestley dual space of F. We will first exhibit an order-homeomorphism ϕ between X and the ordered generalized Cantor space over V . Note that, for any point x ∈ X, the corresponding homomorphism h x : F → 2 must be of the form f¯x for some function f x : V → 2: indeed, h x : F → 2 is equal to f¯x for f x := h x ◦ e. Let us write ϕ for the surjective function from 2V to X that sends f : V → 2 to the element x ∈ X with h x = f¯. We now show that ϕ is an order-embedding, that is, that for any functions f , g : V → 2, we have f  g if, and only if, ϕ( f ) ≤ ϕ(g) in X. First, if ϕ( f ) ≤ ϕ(g) in X, then f¯ ≥ g¯ pointwise, so in particular f = f¯ ◦ e ≥ g¯ ◦ e = g pointwise, which

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.1 Free Distributive Lattices

93

means f  g by definition. For the other direction, note that, if f  g, then the set {u ∈ F | f¯(u) ≥ g(u)} ¯ is a sublattice of F containing V , and must therefore be equal to F by Proposition 4.3. It follows that ϕ is an order-isomorphism between the posets (2V , ) and (X, ≤). We now show that ϕ is a homeomorphism from (2V , π) to (X, τ). Note first that, by Proposition 4.3, the set e[V ] generates F. It follows that the image of e[V ] under  : F → ClpD(X ) generates ClpD(X ). Thus, by definition of the isomorphism (−) the Priestley topology τ on X, to establish continuity of ϕ, it suffices to show that  is clopen for every v ∈ V . Let v ∈ V and f ∈ 2V . We then have that ϕ−1 ( e(v))  if, and only if, f¯(e(v)) = 1, if, and only if, f (v) = 1. Thus, for any ϕ( f ) ∈ e(v) v ∈ V , we have  = v → 1, (4.1) ϕ−1 ( e(v)) which is a clopen set in the product topology π. Since the sets v → 1 and their complements form a subbase for the clopen sets of the topology π, and ϕ is a bijection, we also immediately obtain from the equality (4.1) that ϕ is an open map. Finally, (4.1) shows that the diagram F

 (−)

ϕ−1

e V

ClpD(X )

s

ClpD(2V )

commutes, and therefore the function s is a universal arrow, since the arrow e is  is an isomorphism, and isomorphisms preserve universal arrows universal, ϕ−1 ◦ (−) (see Exercise 4.1.1).  Remark 4.9 The alternative explanation of the free distributive lattice as up-sets instead of down-sets from Example 4.6 carries over to the general case. In particular, note that the function s from Proposition 4.8 may also be described as the function s : V → ClpU(2V ) which sends v to v → 1, where now 2 is equipped with the order in which 0 ≤ 1. A different kind of order symmetry in this context is the fact that the free distributive lattice is anti-isomorphic to itself. In terms of the concrete representation of the free distributive lattice as clopen down-sets of the ordered generalized Cantor space, we may consider the function t : V → ClpU(2V ) op that sends v ∈ V to t(v) := v → 0. Its unique extension t¯ along s is an isomorphism from ClpD(2V ) to ClpU(2V ) op , which may be described concretely by sending a clopen down-set D to the clopen up-set 2V – D.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

94

4 Duality Methods

Remark 4.10 We outline an alternative, more abstract proof of Proposition 4.8, by using some forward references to the categorical language that we will develop in Chapter 5 (see also Exercise 5.3.13). First, using the categorical fact that “left adjoints preserve colimits” (see Exercise 5.3.12), the free distributive lattice over a set V must be the V -fold coproduct of the free distributive lattice over a singleton set. Now, the free distributive lattice on one generator p is easily seen to be the threeelement chain {⊥ < p < }, whose dual is {p, }. Using the fact that the category of Priestley spaces contains arbitrary products of finite posets (see Example 5.35), we may now recover Proposition 4.8 directly from the categorical fact that a dual equivalence sends any coproduct of distributive lattices to a corresponding product of Priestley spaces. The free Boolean algebra over a set V can be defined in an entirely analogous way to the free distributive lattice: it is a Boolean algebra A, together with a function e : V → A, such that, for every function f : V → B with B a Boolean algebra, there exists a unique homomorphism f¯ : A → B such that f¯ ◦ e = f . The proofs of Propositions 4.2 and 4.3 can be carried out in the same way for the free Boolean algebra, as well as the algebraic construction of the free Boolean algebra: this is now the Lindenbaum–Tarski algebra of classical propositional logic. Alternatively, we may combine the free distributive lattice with the Boolean envelope construction (see Section 3.3) to obtain the free Boolean algebra. Lemma 4.11 Let V be a set. The Boolean envelope of the free distributive lattice over V is the free Boolean algebra over V . Proof Let i : V → FDL (V ) be the free distributive lattice over V and let j : FDL (V ) → FDL (V ) − be the Boolean envelope of FDL (V ). We claim that the composite e := j ◦ i : V → FDL (V ) − has the required universal property for the free Boolean algebra. Indeed, for any function f : V → B with B a Boolean algebra, there is first, by definition of the free distributive lattice, a unique lattice homomorphism f¯ : FDL (V ) → B with f¯ ◦ i = f , and then, by definition of the Boolean envelope, a unique homomorphism f¯− : FDL (V ) − → B such that f¯− ◦ j = f¯. We are in the situation shown in the following diagram: j

FDL (V )

FDL (V ) −

f¯ f¯−

i e V

f

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

B

4.1 Free Distributive Lattices

95

and we get f¯− ◦ e = f¯− ◦ j ◦ i = f¯ ◦ i = f . The uniqueness of f¯− is clear from the uniqueness parts of the universal properties of the free distributive lattice and the Boolean envelope.  Corollary 4.12 Let V be a set. The Boolean algebra of clopen sets of the generalized Cantor space, (2V , π), is the free Boolean algebra over V via the function s which sends any v ∈ V to the clopen set v → 1. Proof Combine Proposition 3.34, Proposition 4.8, and Lemma 4.11.



In the case where V is countably infinite, the free Boolean algebra over V is countably infinite, and it is atomless, that is, it does not have any atoms. In fact, one may use model-theoretic techniques to prove that this is, up to isomorphism, the unique countably infinite and atomless Boolean algebra, and it is a central structure in several parts of logic, see, for example, Koppelberg et al. (1989, Corollary 5.16). Its dual is the classical Cantor space 2N (also see Exercise 3.3.9). Exercises for Section 4.1 Exercise 4.1.1 Let V be a set, and suppose that e : V → F, e  : V → F  are functions to distributive lattices F and F , and ϕ : F  → F is a lattice isomorphism such that ϕ ◦ e  = e. Prove that e is a universal arrow if, and only if, e  is a universal arrow. Exercise 4.1.2 How many elements does the free Boolean algebra on two generators have? Draw its Hasse diagram, labeling each of its elements with a corresponding Boolean algebra term. Exercise 4.1.3 Let L be a distributive lattice and suppose that G ⊆ L is a set of generators for L. Prove that there is a surjective homomorphism FDL (G) → L. Conclude that any finitely generated distributive lattice is finite. Formulate and prove the analogous result for Boolean algebras. Note. This exercise is an instance of the general universal algebraic fact that, in a variety of algebras, all finitely generated algebras are finite if, and only if, the free finitely generated algebras are finite. Such varieties are known as locally finite varieties, and the results in this section show in particular that both the varieties of distributive lattices and of Boolean algebras are locally finite. Exercise 4.1.4 This exercise asks you to supply some details of the proof of Proposition 4.4. (a)

Show that ≡ is congruential on T (V ).

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

96

(b)

(c) (d)

4 Duality Methods

Show that T (V )/≡, with operations defined as in the paragraph before Proposition 4.4, satisfies all the defining equations of a distributive lattice. Hint. This is almost immediate from the definition of ≡. Prove in detail that the function f¯ defined in the proof of Proposition 4.4 is a homomorphism and that f¯ ◦ e = f . Prove that if g : T (V )/≡ → L is a homomorphism and g ◦ e = f , then g = f¯.

Exercise 4.1.5 Let V be a set and L a distributive lattice. Show that the poset LV of all functions from V to L with the pointwise order is order isomorphic to the poset of all homomorphisms from FDL (V ) to L also with the pointwise order. Exercise 4.1.6 This exercise outlines a proof of Proposition 4.8 that does not rely on generalities from universal algebra, and gives a more concrete construction of the extension homomorphisms. Let V be a set and (2V , , π) the ordered generalized Cantor space over V . (a) (b)

(c)

Prove that, for each v ∈ V , the set s(v) := v → 1 is a clopen down-set in (2V , , π). Let K be a clopen down-set of (2V , , π). Prove, using the compactness of 2V and the definition of the order, that there exists a finite collection C of 

finite subsets of V such that K = C ∈ C v ∈C s(v). Give a direct proof that, for any function h : V → L, with L a distributive lattice, there exists a unique homomorphism h¯ : ClpD(2V ) → L such that h¯ ◦ s = f .

4.2 Quotients and Subs In many applications of Priestley duality, including some in Chapters 7 and 8 of this book, we use the dual space of a lattice to study its quotient lattices and sublattices, or we use the dual lattice of a space to study its quotient spaces and subspaces. Here, closed subspaces of a Priestley space will play an important role, because they are exactly the subspaces that are themselves Priestley spaces in the inherited order and topology (see Exercise 3.2.4). In this section, we exhibit two Galois connections, between relations and subsets, which in particular allow us to prove that (a) the quotients of a distributive lattice are in an order-reversing bijection with the closed subspaces of its Priestley dual space; and (b) the sublattices of a distributive lattice are in an order-reversing bijection with the Priestley quotients of its Priestley dual space. While (a) and (b) can also be deduced in a more abstract way from Theorem 3.31 using category theory (see Theorem 5.40), the Galois connections that we develop in this chapter do a bit more. In concrete applications,

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

97

these Galois connections are what allow us to define so-called equations on either side of the duality. Let L be a distributive lattice with Priestley dual space X. We will first show how to associate with any binary relation R on L a closed subspace R of X. Here, a pair that belongs to such a binary relation on a lattice is thought of as an equation, and it acts as a constraint defining a subset on the spatial side of the duality. This motivates the following notation and definition. Notation 4.13 When L is a lattice, we will denote, in this subsection, pairs of elements by the notation a ≈ b, instead of (a, b). Definition 4.14 Let L be a distributive lattice with dual Priestley space X. We define the binary relation |= ⊆ X × L 2 by, for any pair of elements a ≈ b ∈ L 2 and x ∈ X, def

x |= a ≈ b ⇐⇒ x ∈  a if, and only if, x ∈  b. For any binary relation R ⊆ L 2 , we define def

x |= R ⇐⇒ for every a ≈ b in R, x |= a ≈ b. Moreover, for any a ≈ b ∈ L 2 , we define a clopen subset a ≈ b of X by a ≈ b := {x ∈ X | x |= a ≈ b} = ( a ∩ b) ∪ ( ac ∩  bc ), and we extend this assignment to any binary relation R on L, by defining R := {x ∈ X | x |= R} = a ≈ b. a≈b ∈R

Conversely, for any element x ∈ X, we define a binary relation ϑ(x) on L by ϑ(x) := ker(h x ) = (Fx × Fx ) ∪ (Ix × Ix ) = {a ≈ b | x |= a ≈ b}. We extend this assignment to subsets S ⊆ X by defining ϑ(S) := ϑ(x) = {a ≈ b | a, b ∈ L such that  a∩S = b ∩ S}. x ∈S

Note that the above definitions of − and ϑ are a special case of the functions u and  introduced in Example 1.16, in the case of the relation |= ⊆ X × L 2 . Proposition 4.15 Let L be a distributive lattice with dual Priestley space X. The two functions − : P (L 2 )  P (X ) : ϑ form a Galois connection, whose fixed points on the left are the lattice congruences on L, and whose fixed points on the right are the closed subspaces of X.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

98

4 Duality Methods

Proof The functions indeed form a Galois connection since they are the contravariant adjoint pair between the powersets of L 2 and X, formed as in Example 1.16, in the case of the relation |=. For the statement about fixed points, we need to show that the image of − consists of the closed subsets of X, and that the image of ϑ consists of the congruences on L. First, since each a ≈ b is clopen, R is closed for every binary relation R. Conversely, let C ⊆ X be any closed set. We show that C = ϑ(C). By adjunction, we have C ⊆ ϑ(C). We prove the other inclusion by contraposition. Let x  C be arbitrary. Using the base for the Priestley topology given in Lemma 3.33, pick a, b ∈ L such that x ∈  a – b and C is disjoint from  a – b. Note that the latter implies that the pair a ≈ a ∧ b is in ϑ(C), since for any y ∈ C, we have y ∈  a if, and only if, y ∈  a ∩ b. However, x ∈  a but x  a ∧ b, so x  ϑ(C), as required. We now show that the image of ϑ consists of the congruences on L. Note that ϑ(S) is a congruence for any S, as it is an intersection of congruences (see Exercise 1.2.16). Let E be a congruence on L. We show that E = ϑ(E). The left-to-right inclusion holds by adjunction. For the other direction, we reason by contraposition, and assume (a, b)  E. Denote by p : L  L/E the lattice quotient by the congruence E. Then, p(a)  p(b), and we assume without loss of generality that p(a)  p(b). By Theorem 3.10 applied to L/E, the filter ↑p(a) and the ideal ↓p(b), we can pick a prime filter G in L/E containing p(a) but not p(b). Then, the set Fx := p−1 (G), the inverse image of G under the homomorphism p, is a prime filter of L which contains a but not b. Moreover, for any (c, d) ∈ E, we have c ∈ Fx if, and only if, p(c) ∈ G, if, and only if, d ∈ Fx , since p(c) = p(d) by assumption. Thus, x ∈ E, and we conclude that (a, b)  ϑ(E).  Remark 4.16 An interesting feature of Proposition 4.15 is that we can start, on either side of the duality, with an unstructured set. Given an arbitrary subset R ⊆ L × L or S ⊆ X, by applying the Galois connection, we always end up with a structured subset, that is, R is not just a subset of X, but it is a closed, hence Priestley, subspace of X, and ϑ(S) is not just a subset of L × L but it is a congruence of L. From Proposition 4.15, we deduce the following theorem, which shows how to explicitly compute the subspace dual to a lattice quotient generated by some equations. Theorem 4.17 Let L be a distributive lattice with dual Priestley space X and let R be a binary relation on L. Then, the lattice congruence generated by R, R, is

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

99

equal to ϑ(R), and the Priestley dual space of L/R is order-homeomorphic to the closed subspace R of X. Proof In general, for any adjunction f : P  Q : g, for any p ∈ P, g f (p) is the minimum of im(g) ∩ ↑p (see Exercise 1.1.8(d) in Chapter 1). In particular, using that im(ϑ) consists of the congruences on L, ϑ(R) is the smallest congruence containing R. Let Y denote the Priestley dual space of the quotient lattice L/R. The dual of the quotient map p : L  L/R is the continuous order-preserving function i : Y → X which can be defined by hi(y) := hy ◦ p, for every y ∈ Y . Note that i is an order embedding: if y  y  in Y , then we can pick b ∈ L/R such that b, and since p is surjective, we can pick a ∈ L such that b = p(a). y∈ b and y    a, so i(y)  i(y  ). Finally, for any x ∈ X, we have that Then, i(y) ∈  a and i(y  )   x ∈ R if, and only if, R ⊆ ϑ({x}) = ker(h x ), using the Galois connection between − and ϑ. The latter holds if, and only if, there exists y ∈ Y such that h x = hy ◦ p. Therefore, the image of i is equal to R = ϑ(R) = R, using the general fact that g f g = g for any adjunction ( f , g), see Exercise 1.1.8(c). Thus, i is a homeomorphism between Y and the closed subspace R, as required.  A second theorem mirrors the previous one, but on the space side. Theorem 4.18 Let X be a Priestley space with dual distributive lattice L and let S be a subset of X. Then, the closure of S, S, is equal to ϑ(S), and the lattice dual to S is isomorphic to the quotient L/ϑ(S). Proof Left as Exercise 4.2.1.



We give a few examples of the duality between lattice quotients and closed subspaces. You are asked to verify some details in Exercises 4.2.3 and 4.2.4. Example 4.19 Let B be the eight-element Boolean algebra; denote its three atoms by a, b, and c, and its dual space by X = {x, y, z}, where Fx = ↑a, Fy = ↑b, and Fz = ↑c. For the equation b ≈ a ∨ b, we get b ≈ a ∨ b = {y, z}, and the generated congruence b ≈ a ∨ b = ϑ(b ≈ a ∨ b) on B is depicted in Figure 4.2. Example 4.20 Let X be the poset depicted on the left in Figure 4.3 and let S = {x, y}. The dual lattice D(X ) and the congruence on D(X ) corresponding to the subspace S ⊆ X are as depicted on the right in the same figure. Furthermore, the subspace S is equal to a ≈ b, where a is the down-set ↓z and b is the down-set ↓x = {x}. Note that the quotient of D (X ) by the congruence a ≈ b is the threeelement chain, which is indeed the lattice dual to the induced poset on the subset S ⊆ X.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

100

4 Duality Methods

a∨b a

b

c

Figure 4.2 A congruence on the eight-element Boolean algebra.

y

z

a

x b

Figure 4.3 A poset, its down-set lattice, and the congruence corresponding to the subset {x, y}. Here, a corresponds to the down-set ↓z and b corresponds to the down-set {x}.

Example 4.21 Recall the Stone–Čech compactification βX of a set X, introduced in Example 3.45. As remarked in Example 3.45, every point of X is isolated in βX, so that X is an open subspace of βX. Recall also from that example that we identify points of X with ultrafilters of P (X ). Note that the set ∗ X := βX – X, known as the remainder of the Stone–Čech compactification βX, is closed. Elements of ∗ X are known as free ultrafilters. Since ∗ X is a closed subspace of the dual space of P (X ), it corresponds to a quotient B of the Boolean algebra P (X ). We will show now that this quotient algebra B is given by the following set of equations on P (X ): {{x} ≈ ∅ | x ∈ X }.

(4.2)

Indeed, if μ is a free ultrafilter and x ∈ X, then ∅  μ and {x}  μ, so μ |= {x} ≈ ∅. On the other hand, if μ is principal, then there is x ∈ X with {x} ∈ μ, see Exercise

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

101

3.3.6. Thus, ∅  μ and {x} ∈ μ, so μ |= {x} ≈ ∅. We have shown that an ultrafilter μ of P (X ) satisfies all the equations in (4.2) if, and only if, μ ∈ ∗ X. Write ϑ for the congruence on P (X ) generated by the set of equations (4.2). Then, ϑ identifies two elements S and T of P (X ) if, and only if, the symmetric difference S + T of S and T is finite. Indeed, for any ultrafilter μ of P (X ), we have μ |= S ≈ T if, and only if, S + T  μ (Exercise 4.2.4). Thus, S and T will be identified by ϑ if, and only if, S + T does not belong to any free ultrafilter, which happens if, and only if, S + T is finite. The Boolean algebra quotient B = P (X )/ϑ, which we identified in this example as the dual algebra of the closed subspace ∗ X of βX, is often denoted P (X )/fin in the literature, and is well studied in set theory and general topology. Sublattices and Priestley Quotient Spaces A very similar story to the one above can be told for sublattices and quotients of Priestley spaces that are themselves Priestley spaces. In fact, the ensuing notion of Priestley space (in)equations is a very important tool in the theory of automata and regular languages, as we will see in Chapter 8. We will now give the necessary definitions and statements of the relevant theorems for a general duality between sublattices and Priestley quotient spaces. We also note that, in fact, the latticequotient–closed-subspace duality from the previous subsection and the sublattice– Priestley-quotient-space duality are, in a sense, dual to each other. We first introduce the notion of quotient space in the context of Priestley spaces. This is an instance of a more general notion of quotient of ordered topological spaces, but we only need it in this setting. Definition 4.22 A preorder  on a Priestley space (X, ≤X , τ) is compatible if ≤X ⊆  and, for any x, y ∈ X, if x  y, then there exists a τ-clopen -down-set K in X such that y ∈ K and x  K. Compatible preorders on X give an intrinsic description of those Priestley spaces Y that are quotients of X, that is, for which there exists a continuous order-preserving surjective map X  Y . We just indicate here what this means, and ask you to fill in the details in the exercises, see Exercises 4.2.5 and 4.2.6. First, for any p : X  Y , the preorder  p on X defined by x  p x  if, and only if, p(x) ≤Y p(x  ) is compatible. Conversely, if  is a compatible preorder on X, denote by Y the poset reflection of the preordered set (X, ), as defined in Exercise 1.1.5. That is, as a set, Y is X/≡ where ≡ :=  ∩ !. Let us write q : X  Y for the quotient map. Then, the partial order ≤Y on the poset reflection is defined, for any y = q(x) and y  = q(x  ) in Y , by y ≤Y y  if, and only if, x  x . The quotient topology on Y is defined by τY := {U ⊆ Y | q−1 (U) ∈ τX }. Then, the ordered topological

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

102

4 Duality Methods

space (Y, ≤Y , τY ) is a Priestley space, and the function q is a continuous orderpreserving map. Moreover, for any Priestley space (Z, ≤ Z , τZ ), any continuous order-preserving map f : (X, , τX ) → (Z, ≤ Z , τZ ) factors uniquely through the map q. We will say that (Y, ≤Y , τY ) is the quotient of the Priestley space X by the compatible preorder  and will denote it by X/. Example 4.23 Consider a Priestley space (X, τ, ≤) with dual lattice L. Notice (see Exercise 3.3.2) that then (X, τ, =) is also a Priestley space. In fact, the dual lattice of this Priestley space is the Boolean algebra L − of all clopen subsets of X, also known as the Boolean envelope of L, see Corollary 3.41. Of course, L is a sublattice of L − . The dual of the inclusion of L into L − is simply the identity map on X, viewed as a continuous order-preserving map idX : (X, τ, =) → (X, τ, ≤). The corresponding compatible preorder is the partial order ≤ of (X, τ, ≤). More generally, any injective homomorphism i : L → B, with B a Boolean algebra, factors as the composition of e : L → L − and i¯ : L − → B. Dually, denoting by (Y, π) the Boolean space dual to B, this gives a quotient map of Boolean spaces f : (Y, π) → (X, τ), followed by idX : (X, τ, =) → (X, τ, ≤). This shows that sublattices of Boolean algebras can be understood dually by a Boolean equivalence relation together with a Priestley order on the quotient. As before, the correspondence between compatible preorders and sublattices allows us to view pairs of elements (x, y) from X as constraints on L yielding sublattices. However, as Example 4.23 already shows, equating elements of the space is not a fine enough notion to witness all sublattices. In fact, as we will see, spatial equations witness Boolean subalgebras, whereas we will need to think of pairs (x, y) as spatial inequations in order to be able to witness all sublattices. For this reason, we introduce the following notation. Notation 4.24 Let X be a set. We denote by I (X ) the set of all formal inequations x  y where x, y ∈ X. That is, I (X ) = {x  y | x, y ∈ X }. Given a set of inequations T ⊆ I (X ), we call RT := {(x, y) | x  y ∈ T } the binary relation corresponding to T. Notice that, up to isomorphism, I (X ) is just X 2 , and subsets T ⊆ I (X ) are just binary relations on X. However, when studying quotients of spaces, we need to be a bit more careful with our notation than in the analogous setting of Notation 4.13 for quotients of lattices, because we here need to carefully distinguish whether we interpret a pair (x, y) of points in a Priestley space as an inequation or as an equation. We begin with the first interpretation. For the second interpretation, see Notation 4.29.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

103

Definition 4.25 Let L be a distributive lattice with dual Priestley space X. We define the binary relation |= ⊆ L × I (X ) by, for any element a ∈ L and x, y ∈ X, def

a |= x  y ⇐⇒ y ∈  a implies x ∈  a. When a |= x  y, we will say that a satisfies the (spatial) inequation x  y. Similarly, for a subset A ⊆ L, we write A |= x  y and say that A satisfies x  y provided every a ∈ A satisfies x  y. This relation allows us to define a map − : P (L) → P (X 2 ), by, for a ∈ L and x, y ∈ X, def

x a y ⇐⇒ a |= x  y  and for any subset A ⊆ L,  A := a ∈ A a , that is, def

x  A y ⇐⇒ A |= x  y. We obtain a map −I : P (X 2 ) → P (L), by, for x, y ∈ X, (x, y)I := Iy ∪ Fx = {a ∈ L | a |= x  y}, and for any R ⊆ X 2 , we define the subset RI of L by RI := (x, y)I = {a ∈ L | for all (x, y) ∈ R, a |= x  y}. (x, y) ∈R

In order to be able to talk about inequations and sets of inequations directly, we will also use, for any (x, y) ∈ X 2 , the notation x  y for (x, y)I , and, for any T ⊆ I (X ), the notation T  for the set RT I , where RT is the binary relation corresponding to T. Proposition 4.26 Let L be a distributive lattice with dual Priestley space X. The two functions − : P (L)  P (X 2 ) : −I form a Galois connection, whose fixed points on the left are the sublattices of L, and whose fixed points on the right are the compatible preorders on X. Proof The functions indeed form a Galois connection since they are the contravariant adjoint pair between the powersets of L and X 2 , formed as in Example 1.16, in the case of the relation |=. Given x, y ∈ X, it is easy to verify that the set x  y is a sublattice of L (see Exercise 4.2.8(a), and thus so is RI , being an intersection of sublattices, for any R ⊆ X 2 . Now let M be a sublattice of L. We want to show M =  M I . By adjunction, we have M ⊆  M I . For the converse inclusion, let a ∈  M I be a, we have a |= x  y, so x  M y, which arbitrary. Note that, for any (x, y) ∈  ac × 

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

104

4 Duality Methods

c  means that we can pick a (x,y) ∈ M with x ∈ a (x,y) and y ∈ a (x,y) . Now, for any fixed x ∈  a c , we get   a⊆ { a a }. (x,y) | y ∈ 

Applying compactness, the union of a finite subcover yields an element bx ∈ M  bx . Now,  a = { bx | x ∈  a c } and, again by compactness, with a ≤ bx and x   there is a finite set F ⊆  a c with   bx . a= 

x ∈F

That is, a = x ∈F bx, which is an element of M. Given a ∈ L, the relation a = ( a × X ) ∪ (X ×  a c ) is clearly a preorder which ac ×  a. That is, contains ≤X . Also, it is compatible since, if x a x , then (x, x  ) ∈   a while x   a and the set  a is a a -down-set (see Exercise 4.2.8(b)). Further, x ∈ for A ⊆ L,  A is again a compatible preorder since these are closed under arbitrary intersections (see Exercise 4.2.8(c)). Now, to see that all compatible preorders are fixed points of the Galois collection, let  be a compatible preorder on X. Again, we have  ⊆   I by adjunction. For the reverse inclusion, suppose x  y. Since  is compatible, we have a clopen V ⊆ X with y ∈ V , x  V , and V is a -down-set. Since ≤X ⊆ , V is also a a. The ≤X -down-set, and thus, since V is also clopen, we can pick a ∈ L with V =  a fact that  a is a -down-set precisely says that a ∈ I . Also, the fact that y ∈   and x   a means that x a y and thus (x, y)    I , as required. In the same way as for quotient–subspace duality, we now get the following duality between sublattices and compatible preorders on its dual space; compare Theorems 4.17 and 4.18. Theorem 4.27 (Schmid, 2002) Let L be a distributive lattice with dual Priestley space X. (a)

(b)

For any subset A of L, the sublattice generated by A is equal to  A I , and the Priestley dual space of this sublattice is order-homeomorphic to the Priestley quotient X/ A. For any binary relation R on X, the smallest compatible preorder containing R is equal to R I , and the lattice dual to the Priestley quotient by this preorder is isomorphic to RI .

Theorem 4.27 allows one to characterize sublattices of a lattice L by sets of inequations over the dual space of L. As we will see in Section 8.3, this is a method used in automata theory to prove the decidability of the membership problem for classes of regular languages. Apart from characterizing sublattices of a lattice L,

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

105

Theorem 4.27 can also be used to separate sublattices, as identified in the following corollary. Corollary 4.28 Let L be a distributive lattice with dual Priestley space X and let M1 and M2 be two sublattices of L. Then, M2  M1 if, and only if, there exists an inequation x  y over X and a ∈ M2 so that a |= x  y but M1 |= x  y. We now spell out the specialization of the subalgebra–quotient-space duality to the Boolean case. Recall from Proposition 3.36 that a Priestley space corresponds to a Boolean algebra if, and only if, the order of the space is trivial. It follows that a compatible preorder  yields a Boolean quotient if, and only if,  is an equivalence relation or, equivalently, if it is symmetric. We call symmetric compatible preorders compatible equivalence relations, and, in situations where we are interested in Boolean subalgebras, we use the stronger notion of spatial equation that we define now. Notation 4.29 For any set X, we denote by E(X ) the set of all formal equations x ≡ y where x, y ∈ X. That is, E(X ) = {x ≡ y | x, y ∈ X }. As before, when T ⊆ E(X ) is a set of formal equations, we denote by RT ⊆ X 2 the corresponding binary relation on X. Analogously to what we did above, we define a notion of satisfaction for formal equations on a space. Definition 4.30 Let B be a Boolean algebra and X its Priestley dual space. We define the binary relation |= ⊆ B × E(X ) by, for any element a ∈ B and x, y ∈ X, a |= x ≡ y ⇐⇒ x ∈  a if, and only if, y ∈  a. When a |= x ≡ y, we will say that a satisfies the (spatial) equation x ≡ y. Similarly, for a subset A ⊆ B, we write A |= x ≡ y and say that A satisfies x ≡ y provided every a ∈ A satisfies x ≡ y. This relation |= allows us to define a map ≡− : P (B) → P (X 2 ), by, for a ∈ B and x, y ∈ X, x ≡a y ⇐⇒ a |= x ≡ y,  and for any subset A ⊆ B, ≡ A := a ∈ A ≡a , that is, x ≡ A y ⇐⇒ A |= x ≡ y. Similarly we obtain a map −E : P (X 2 ) → P (B), by, for x, y ∈ X, a, if, and only if, y ∈  a }, (x, y)E := (Ix ∩ Iy ) ∪ (Fx ∩ Fy ) = {a ∈ B | x ∈ 

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

106

4 Duality Methods

and for any R ⊆ X 2 , we define the subset RE of B by RE := (x, y)E = {a ∈ B | for all (x, y) ∈ R, a |= x ≡ y}. (x,y) ∈R

In order to be able to talk about equations directly, we will also use the notation x ≡ y for (x, y)E and, for any subset T ⊆ E(X ), we write T  for RT E , where RT is the binary relation corresponding to T. Proposition 4.31 Let B be a Boolean algebra with dual Priestley space X. The two functions ≡− : P (B)  P (X 2 ) : −E form a Galois connection, whose fixed points on the left are the Boolean subalgebras of B, and whose fixed points on the right are the compatible equivalence relations on X. Proof See Exercise 4.2.9.



Corollary 4.32 Let B be a Boolean algebra with dual space X. (a)

(b)

For any subset A of B, the Boolean subalgebra generated by A is equal to ≡ A E , and the dual space of this subalgebra is homeomorphic to the quotient space X/≡ A. For any binary relation R on X, the smallest compatible equivalence relation containing R is equal to ≡R E , and the Boolean algebra dual to the quotient by this equivalence relation is isomorphic to RE .

Proof Deduced from Proposition 4.31 in a similar way to Theorems 4.17 and 4.18.  Note that, in part (a) of Corollary 4.32, we could have used −I instead of −E since, when applying −I to a compatible equivalence relation, the corresponding lattice is a Boolean algebra. Further, ≡ A I and ≡ A E are also equal to  A E . Similarly, in part (b) of Corollary 4.32, since the lattice RI generates the Boolean algebra RE , the same compatible equivalence relation corresponds to both of these subsets of B. Finally, ≡R I and ≡R E are both equal to R E , since an inequality is satisfied by a Boolean subalgebra if, and only if, its opposite is also. Remark 4.33 One problem with compatible preorders, which has sometimes hampered their successful application, is that it is difficult to understand intrinsically in a Priestley space how to obtain the compatible preorder generated by a binary relation on the space. The existence of the Galois connection of Proposition 4.26 frees us from this problem and allows us to specify sublattices from arbitrary sets E of inequality constraints of the form x  y. This is similar to the fact that, in logic, we do not need to identify the full theory of a class of structures that we are interested in, since we may be able to capture it by a much smaller set of axioms.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

107

As mentioned above, we will see dual space (in)equations in action in Chapter 8. Below, we will already give two elementary examples. Example 4.34 (Sublattices of the free distributive lattice) Coming back to the example of the free distributive lattice on two generators, Example 4.6, consider the sublattices L 1 = {⊥, p, } and L 2 = {⊥, p ∨ q, }. One may prove from the definitions that the corresponding quotients of 2 {p,q } are given by the equivalence relations ≡1 and ≡2 , where ≡1 has classes {00, 01} and {10, 11}, while ≡2 has classes {00} and {01, 10, 11}. We leave it as Exercise 4.2.11 to classify the other sublattices of this lattice. Example 4.35 (Equations for a subalgebra) In this example we will see how we can use extraneous structure on a dual space and lattice to identify a smaller set of equations for a subalgebra of an infinite Boolean algebra; this technique is further exploited in Section 8.3. Consider the set Z of integers, and denote by Z+ the subset of positive integers and by Z− the subset of negative integers; so Z = Z− ∪ {0} ∪ Z+ . Let M be the Boolean subalgebra of P (Z) consisting of all those subsets S of Z such that both S ∩ Z+ is either finite or co-finite, and S ∩ Z− is either finite or co-finite. One may then show (see Exercise 4.2.12) that the dual space of M is the “two-point compactification of Z” Z+∞ −∞ := Z ∪ {−∞, +∞}, which topologically is the disjoint union of the one-point compactification Z+ ∪ {+∞} of Z+ with the discrete topology, the one-point compactification Z− ∪ {−∞} of Z− with the discrete topology, and the one point space {0}. (For the one-point compactification, see Example 3.46 and Exercise 3.2.2.) Note that, since Z is the disjoint union of Z+ , Z− , and {0}, every free ultrafilter of P (Z) contains exactly one of Z+ or Z− , because a free ultrafilter clearly cannot contain {0}. The dual of the inclusion M → P (Z) is the surjective function β(Z)  Z+∞ −∞, ⎧ k ⎪ ⎪ ⎪ ⎨ μ → ⎪ +∞ ⎪ ⎪ −∞ ⎩

if if if

{k} ∈ μ where k ∈ Z, μ free and Z+ ∈ μ, μ free and Z− ∈ μ.

Thus, the compatible preorder on βZ corresponding to the subalgebra M of P (Z) is the equivalence relation in which each k ∈ Z is only related to itself, and two free ultrafilters μ and ν are related provided they either both contain Z+ , or both contain Z− . That is, the remainder is split into two uncountable equivalence classes and each free ultrafilter is related to uncountably many other free ultrafilters.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

108

4 Duality Methods

By contrast, we will now show that, by using the successor structure on Z, the subalgebra M can be “axiomatized” by a much “thinner” set of equations. The successor function Z → Z, which sends k ∈ Z to k + 1, has as its discrete dual the complete homomorphism P (Z) → P (Z) which sends a subset S ∈ P (Z) to the set S − 1 := {k ∈ Z | k + 1 ∈ S} = {s − 1 | s ∈ S} . For μ ∈ βZ, write μ + 1 := {S ∈ P (Z) | S − 1 ∈ μ} = {S + 1 | S ∈ μ}. This is a well-defined function βZ → βZ, as it is the dual function of the homomorphism S → S − 1 on P (Z); it is also the unique continuous extension of the successor function Z → Z, when we view the codomain as a subset of βZ. To describe the equational basis for the sublattice M of P (Z), consider the set of equations μ + 1 ≈ μ, as μ ranges over ∗ Z, where we recall that ∗ Z := βZ – Z, the remainder of βZ, see Example 4.21. We will show that the sublattice M of P (Z ) contains exactly those S ∈ P (Z) that satisfy all of these equations, that is, we will prove that M

=

 μ + 1 ≈ μ | μ ∈ ∗ Z .

(4.3)

To this end, note first that, for S ∈ P (Z), S satisfies μ + 1 ≈ μ if, and only if, both μ and μ + 1 contain S, or neither μ nor μ + 1 contains S. Now, for the left-to-right inclusion of (4.3), let S ∈ M and let μ be a free ultrafilter of P (Z). We show that S |= μ + 1 ≈ μ. Since μ is prime and Z+ ∪ (Z – Z+ ) = Z ∈ μ, it follows that either Z+ ∈ μ or Z – Z+ ∈ μ. We treat the case Z+ ∈ μ and leave the other as an exercise. Since S ∈ M, we have that S ∩ Z+ is either finite or co-finite. If S ∩ Z+ is finite, then, as μ is free, S  μ. Also S ∩ Z+ finite implies that (S − 1) ∩ Z+ is finite and thus S − 1  μ. So S |= μ + 1 ≈ μ. If, on the other hand, S ∩ Z+ is co-finite, then, as (Z+ – S) ∪ (S ∩ Z+ ) = Z+ ∈ μ, and μ is free, it follows that S ∩ Z+ ∈ μ. Furthermore, S ∩ Z+ co-finite implies that (S − 1) ∩ Z+ is also co-finite and by the same argument we have S − 1 ∈ μ so that S |= μ + 1 ≈ μ. For the right-to-left inclusion of (4.3), we reason contrapositively, and suppose S  M. Then, S ∩ Z+ is neither finite nor co-finite, or S ∩ Z− is neither finite nor co-finite. Again, we treat the first case and leave the second as an exercise. If S ∩ Z+ is neither finite nor co-finite it follows that there is an infinite set T ⊆ Z+ such that, for each k ∈ T, kS

but

k + 1 ∈ S.

By the prime filter-ideal theorem, Theorem 3.10, here applied to the Boolean algebra P (Z), pick an ultrafilter μ of P (Z) which contains the filter ↑T and is disjoint from the ideal I consisting of all finite subsets of P (Z). Since μ is disjoint from I, it is free, and since ↑T ⊆ μ we have T ∈ μ. Now as S ∩ T = ∅ it follows that S  μ and since T + 1 ⊆ S, or equivalently, T ⊆ S − 1, it follows that S − 1 ∈ μ. That is,

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

109

we have exhibited a free ultrafilter μ such that S |= μ + 1 ≈ μ and this completes the proof of (4.3). We note that Example 4.35 is related to a well-known language from descriptive complexity theory. The monoid of integers under addition is the so-called syntactic monoid of the language called “majority,” consisting of all bitstrings with a majority of 1s; see Exercise 8.3.6. In Chapter 8 we will also need the duality between complete Boolean subalgebras of a powerset algebra and equivalence relations on the underlying set. This is a dicrete version of the duality between spatial equations and Boolean subalgebras stated in Proposition 4.31. Like all other correspondences in this section, it is also possible to treat this duality via the Galois connection obtained from a relation |= ⊆ P (S) × P (S × S) defined by a |= s ≡ s  if, and only if, s ∈ a ⇐⇒ s  ∈ a, but we here give an alternative, direct proof, in Theorem 4.36. Recall that a complete Boolean subalgebra of a complete Boolean algebra B is a Boolean subalgebra A such that, for any S ⊆ A, the supremum and the infimum of S in B are in A. Note that, using negation, it suffices to assume this for suprema or for all infima. Also note that for A to be a complete Boolean subalgebra it is not sufficient to merely require that A be a Boolean algebra which is complete in its order (see Exercise 1.2.6(a)). In Theorem 4.36, we will consider the poset Subc (B) of complete Boolean subalgebras ordered by inclusion, and the poset Eq(S) of equivalence relations on a set S, also ordered by inclusion. Both are complete lattices. For an equivalence relation ≡ on a set S, we will call a subset u of S invariant provided u |= s ≡ s  for all s, s  ∈ S and we write ≡ for the collection of invariant subsets of S. Theorem 4.36 Let S be a set and let B := P (S) be its Boolean algebra of subsets. For any equivalence relation ≡ on S, the set ≡ is a complete Boolean subalgebra of B. Moreover, the assignment ≡ → ≡ is an anti-isomorphism between Eq(S) and Subc (B). Proof Let f : S → S/≡ be the quotient function associated to ≡. Note that, essentially by definition, ≡ is exactly the image of the dual complete homomorphism f −1 : P (S/≡) → B, which is injective and the upper adjoint of the forward image function f [−] : P (S) → P (S/≡), see Exercise 1.3.10. Conversely, given a complete Boolean subalgebra A of B, the inclusion i : A → B is an injective complete homomorphism and thus, it has a lower adjoint i  : B → A which is surjective. As in the finite case, see the proof of Lemma 1.23, i  sends completely join irreducibles to completely join irredicibles, and thus atoms to atoms. That is, i  restricts to a function f : B → At( A). Note that since i  preserves arbitrary joins, i  is, up to isomorphism, the forward image function f [−]. Therefore, i  is surjective if, and

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

110

4 Duality Methods

only if, f is, and thus f is the quotient function associated to a uniquely determined equivalence relation on S. Clearly these two assignments are order reversing and inverse to each other.  Exercises for Section 4.2 Exercise 4.2.1 Prove Theorem 4.18. Exercise 4.2.2 Consider the unique homomorphism h : FDL (p, q) → FBA (p) that sends p to p and q to ¬p. Show that h is surjective, and compute the dual injective function. Exercise 4.2.3 Prove the statements made in Examples 4.19 and 4.20. Exercise 4.2.4 Let ϑ be a congruence on a Boolean algebra B. (a) (b) (c)

Prove that, for any a, b ∈ B, a ϑ b if, and only if, a + b ϑ 0. Here, a + b denotes the symmetric difference of a and b, that is, a + b := (a ∧ ¬b) ∨ (¬a ∧ b). Prove that, for any ultrafilter μ of B, we have μ |= a ≈ b if, and only if, a + b  μ. Conclude that, for any binary relation R on B, we have x ∈ R if, and only if, x  a + b for all (a, b) ∈ R.

Exercise 4.2.5 Let p : X → Y be a continuous order-preserving map between Priestley spaces. Prove that the relation  := p−1 (≤Y ) on X is a compatible preorder. Exercise 4.2.6 Let (X, τX , ≤X ) be a Priestley space and let  be a compatible preorder on X. Define ≡ :=  ∩ !, let Y := X/≡, and denote by q : X  Y the quotient map. (a) (b) (c) (d) (e)

Show that, for any x, x  ∈ X, x ≤X x  implies x  x . Prove that the relation ≤Y on Y defined, for y = q(x) and y  = q(x  ) in Y , by y ≤Y y  if, and only if, x  x , is a well-defined partial order. Prove that, with the quotient topology τY := {U ⊆ Y | q−1 (U) ∈ τX }, (Y, τY , ≤Y ) is a Priestley space. Prove that q : X → Y is continuous and order preserving. Prove that, for any Priestley space Z and any continuous f : X → Z such that x  x  implies f (x) ≤ Z f (x  ), there exists a unique continuous orderpreserving f¯ : Y → Z such that f = f¯ ◦ q.

Exercise 4.2.7 Prove the assertions made in Example 4.23. Exercise 4.2.8 Let L be a distributive lattice and X its Priestley dual space. (a)

For x, x  ∈ X, show that x  x   is a sublattice of L.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.2 Quotients and Subs

(b)

(c)

111

For each a ∈ L, show that a is a compatible preorder on X, that the equivalence classes of the corresponding equivalence relation are  a and  ac , c and that  a≤ a is the only non-trivial relation in the quotient order. Conclude that a I is the three-element sublattice {0, a, 1} of L. Show that compatible preorders on X are closed under arbitrary intersections.

Exercise 4.2.9 (a) Verify that the functions in Proposition 4.31 constitute the Galois connection given by the relation a |= x ≡ y. (b) Verify that, for any R ⊆ X 2 , RE is closed under complementation and is thus a Boolean subalgebra of B. (c) Prove that, for any subset A ⊆ B, ≡ A =  A , where A = A ∪ {¬a | a ∈ A} and conclude that it is a compatible equivalence relation on X. (d) Using the above and Proposition 4.26 complete the proof of Proposition 4.31 and its Corollary 4.32. (e) State and prove results analogous to Proposition 4.31 and Corollary 4.32 for the discrete duality between complete and atomic Boolean algebras and sets. Exercise 4.2.10 Let ϑ be an equivalence relation on a Boolean space X. Prove that the following conditions are equivalent: (i) (ii) (iii)

ϑ is clopen as a subset of X × X; [x]ϑ is clopen for every x ∈ X; and ϑ has finitely many classes and the corresponding quotient map X → X/ϑ is continuous (where X/ϑ is a finite discrete space).

Hint. (i) =⇒ (ii): Show that both [x]ϑ = π2 [ϑ ∩ ({x} × X )] and π2 [ϑ c ∩ ({x} × X )] are closed. (ii) =⇒ (iii): Use compactness of X. (iii) =⇒ (i): Consider the product map X × X → X/ϑ × X/ϑ and the inverse image of the diagonal in X/ϑ × X/ϑ. Exercise 4.2.11 Building on Example 4.34, identify the sublattices of the free distributive lattice on two generators and the corresponding compatible quasi-orders on 2 {p,q } . Exercise 4.2.12 Let M be the Boolean subalgebra of P (Z) consisting of all those S ⊆ Z such that both S ∩ Z+ is either finite or co-finite and S ∩ Z− is either finite or co-finite, see Example 4.35. (a) (b)

Show that M  M − × P ({0}) × M + , where M − and M + are the Boolean algebras of all finite or co-finite subsets of Z− and Z+ , respectively. Show that the dual space of M is the topological sum (i.e., disjoint union) of the one-point compactification Z− ∪ {−∞} of Z− , the one-point compactification Z+ ∪ {+∞} of Z+ , and the one-point space {0}.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

112

4 Duality Methods

Exercise 4.2.13 Let X be a set and βX the dual space of P (X ). (a)

Show that there is a one-to-one correspondence between each of (1) (2) (3)

(b)

the Boolean subalgebras B of P (X ); the continuous surjections f : βX → Y with Y a Boolean space; and the set functions h : X → Y with dense image.

In particular, show that for each L ⊆ X, we have βX  L βX = L

and, for a subalgebra B and corresponding continuous surjection f and set function with dense image h, we have Y

L ∈ B ⇐⇒ f [ L βX ] is open in Y ⇐⇒ h[L] is open in Y, and in this case we have Y  L βX ] = h[L] , LY = f [

where  ( ) is the Stone map, ( ) is topological closure, and the decorations refer to the ambient space in question. 4.3 Unary Operators In all the dualities discussed in this book so far, the morphisms on the algebraic side have been the homomorphisms, that is, the maps preserving all of the lattice structure. In this section, we will relax this condition and study maps between distributive lattices that only preserve finite meets, but not necessarily finite joins. Such functions are also known as unary normal multiplicative operators in the literature. The dualities developed in this section originate with the seminal works Jónsson and Tarski (1951, 1952); also see the notes at the end of this chapter. We show that there is still a dual equivalence of categories if we generalize, on the space side, from continuous order-preserving functions to certain relations that are compatible with the order and topology of the Priestley space. We begin by introducing some necessary notation and defining what this means precisely. Notation 4.37 Here, and in what follows, we use some common notations for composition, forward and inverse image for binary relations: Let R ⊆ X × Y and S ⊆ Y × Z be binary relations. We often use infix notation, writing x Ry for (x, y) ∈ R. The composition R · S is the relation from X to Z defined by R · S := {(x, z) | there exists y ∈ Y such that x RySz} . Note that this left-to-right notation for relational composition differs from the right-to-left notation ◦ that we use for functional composition (Notation 1.3). That

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.3 Unary Operators

113

is, if f : X → Y and g : Y → Z are functions with graphs R f ⊆ X × Y and Rg ⊆ Y × Z, respectively, then the relational composition R f · Rg is the graph of the functional composition g ◦ f . The converse of R is the relation R−1 := {(y, x) ∈ Y × X : x Ry}. For any subset U of X, we write R[U] for the relational direct image, that is, R[U] := {y ∈ Y | there exists u ∈ U such that uRy} . The same set is sometimes denoted ∃R [U]. For singleton subsets {x} of X, we write R[x] instead of R[{x}]. Converse relations allow us, in particular, to define relational inverse image R−1 [V ] for any V ⊆ Y , by taking the direct image of the converse relation, that is, R−1 [V ] = {x ∈ X | there exists v ∈ V such that x Rv} . Finally, we also use the relational universal image, ∀R , defined, for any U ⊆ X, by ∀R [U] := {y ∈ Y | for all x ∈ X, if x Ry, then x ∈ U},

(4.4)

and we note that ∀R [U] = {y ∈ Y | R−1 [y] ⊆ U} . For later use, we note a convenient formula for switching between direct and universal relational image: ∀R [U] = Y – R[X – U],

for any U ⊆ X .

(4.5)

The import of the operation ∀R introduced in (4.4) stems from the fact that if R ⊆ X × Y is any relation between sets, then the function R−1 [−] : P (Y ) → P (X ) preserves arbitrary joins, and the function ∀R [−] : P (X ) → P (Y ) is its upper adjoint (see Exercise 4.3.3(a)). Furthermore, we will now define the property of upward order-compatibility for a relation R between Priestley spaces, which turns out to be equivalent to the fact that the operations R−1 and ∀R restrict correctly to the sublattices D(X ) and D(Y ) (see Exercise 4.3.3(b)). As we will see in the proof of Proposition 4.39, these basic order-theoretic facts, combined with two topological continuity properties for the relation R, are fundamental to the duality for unary operators. Note that the notion of “compatibility” for an arbitrary relation between Priestley spaces, which we introduce now, is distinct from the notions of compatibility for preorders and equivalence relations that we saw in Section 4.2. Definition 4.38 Let X and Y be Priestley spaces and let R ⊆ X × Y be a relation. We say that R is: • upward order-compatible if ≥ · R · ≥ ⊆ R, that is, for any x, x  ∈ X and y, y  ∈ Y , whenever x  ≥ x Ry ≥ y , we have x  Ry ;

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

114

4 Duality Methods

• upper Priestley continuous if, for every clopen up-set K ⊆ Y , the set R−1 [K] is clopen; • point-closed if, for every x ∈ X, the set R[x] is closed; and • upward Priestley compatible if R is upward order-compatible, upper Priestley continuous, and point-closed. The first aim of this section is to prove (Proposition 4.39) that the finite-meetpreserving functions between two distributive lattices are in one-to-one correspondence with the upward Priestley compatible relations between their respective dual spaces. This correspondence will generalize the correspondence between homomorphisms and continuous order-preserving functions of Priestley duality (see Exercise 4.3.4). It also yields a new duality theorem, as we will remark at the end of this section and prove in Chapter 5 once we have the appropriate categorical terminology in place. We will also show at the end of this section how to obtain an analogous correspondence between finite-join-preserving functions and downward Priestley compatible relations.

The Finite Case To motivate our proof of the general case (Proposition 4.39), let us proceed as we did in Chapters 1 and 3 and first examine the finite case. Let L and M be finite distributive lattices. In Chapter 1, we saw that every homomorphism h : M → L uniquely arises as f −1 for some order-preserving f : J (L) → J (M). This function f was obtained as the restriction to J (L) of the lower adjoint of h. Crucially, we showed in Lemma 1.23 that this lower adjoint sends any element of J (L) to J (M). This way of defining a function dual to h first uses that h preserves finite meets for the existence of the lower adjoint f , and then uses that h preserves finite joins to show that f sends join-primes to join-primes. Now, when we consider functions between finite distributive lattices that only preserve meets, but not necessarily joins, the lower adjoint still exists, but it may no longer restrict correctly to join-prime elements. Instead, for a meet-preserving function h : M → L with lower adjoint f : L → M, recall from Proposition 1.21 that every element a ∈ L is a finite join of join-prime elements. Therefore, since f : L → M preserves finite joins and J (L) join-generates L, the function f is uniquely determined by its restriction to join-prime elements: for any a ∈ L, we  have f (a) = f [↓a ∩ J (L)]. Moreover, since J (M) join-generates M, in order to dually encode the function f , and therefore h, it suffices to know the value f (p) for each p ∈ J (L). To this end, we define R := {(p, q) ∈ J (L) × J (M) | q ≤ f (p)},

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

(4.6)

4.3 Unary Operators

115



so that f (p) = R[p] for every p ∈ J (L). Note that we can express R in terms of h using the adjointness. Indeed, for any p ∈ J (L) and q ∈ J (M), we have q ≤ f (p) ⇐⇒ ∀b ∈ M ( f (p) ≤ b =⇒ q ≤ b) ⇐⇒ ∀b ∈ M (p ≤ h(b) =⇒ q ≤ b)  =⇒ q ∈  ⇐⇒ ∀b ∈ M (p ∈ h(b) b),  denotes the lattice isomorphism between a finite distributive where we recall that (−) lattice and the down-set lattice of its poset of join-prime elements (Proposition 1.21). We call the binary relation R ⊆ J (L) × J (M) defined by (4.6) the relation dual to h. This relation R is upward order-compatible, see Exercise 4.3.1. Moreover, if we view the finite posets J (L) and J (M) as Priestley spaces by equipping them with the discrete topology, then any relation between them is trivially upper Priestley continuous and point-closed. The original meet-preserving function h : M → L can be recovered from R, using the following equality, which holds for any b ∈ M:  = ∀R−1 [ b], h(b)

(4.7)

where we recall that ∀R is the universal image defined in (4.4). Writing out the definitions, (4.7) expresses the fact that, for any p ∈ J (L), p ≤ h(b) ⇐⇒ for all q ∈ J (M), if pRq, then p ≤ b, which can be proved using the lower adjoint f of h and the definition of R (see Exercise 4.3.1). This concludes our informal description, in the case of finite distributive lattices, of the duality between meet-preserving functions and upward Priestley compatible relations. Summing up, we have represented the meet-preserving function h by an upward order-compatible relation R between the dual posets J (L) and J (M), from which h can be recovered as the upper adjoint of the unique join-preserving  function f which is defined for p ∈ J (L) by f (p) := R[p]. As an instructive exercise (see Exercise 4.3.1), we invite you to verify the claims that were left unproved here, although they are also direct consequences of the general duality theorem that we prove below (see Exercise 4.3.3 for details about the relationship). A reader familiar with modal logic may have recognized in the equation (4.7) the definition of the  (“box”) operator associated to a Kripke relation R; more on this in Section 4.4.

The General Case In the remainder of this section, we generalize the ideas outlined above to finitemeet-preserving functions h between arbitrary distributive lattices that are not

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

116

4 Duality Methods

necessarily finite. As in Chapter 3, join-primes have to be replaced by points of the dual space, but the underlying ideas are the same as in the finite case. Proposition 4.39 Let L and M be distributive lattices with Priestley dual spaces X and Y , respectively. For any finite-meet-preserving h : M → L, there exists a unique upward Priestley compatible relation, R ⊆ X × Y , such that,  = ∀ R−1 [ b]. for any b ∈ M, h(b) This relation R may be defined explicitly, for x ∈ X, by  R[x] := { b | b ∈ M, x ∈ h(b)},

(4.8)

(4.9)

or, equivalently, R := {(x, y) ∈ X × Y | for all b ∈ M, if h(b) ∈ Fx, then b ∈ Fy }.

(4.10)

Proof Let R ⊆ X × Y denote the relation defined in (4.9) and (4.10). We will now establish three properties, which suffice to conclude: (a) (b) (c)

R satisfies (4.8); R is upward Priestley compatible; and R is the unique relation with properties (a) and (b).

For (a), unfolding the definitions, we need to prove that, for any x ∈ X and b ∈ M,  ⇐⇒ ∀y ∈ Y, if x Ry, then y ∈  x ∈ h(b) b.

(4.11)

The left-to-right direction is clear by definition of R. For the converse, we rea this means that the homomorphism son contrapositively. Suppose that x  h(b); h x : L → 2 sends h(b) to 0. Since h is finite-meet-preserving, the composite function k := h x ◦ h : M → 2 is finite-meet-preserving. Thus, the set F := k −1 (1) is a filter which does not contain b. By the prime filter theorem (Theorem 3.10), pick a prime filter Fy containing F and still not containing b. The fact that F ⊆ Fy is easily seen to be equivalent to x Ry, while y   b, establishing that the right-hand side of (4.11) fails, as required. For (b), we first show that (a) already yields that R is upper Priestley continuous. Indeed, any clopen up-set K ⊆ Y is equal to Y –  b for some b ∈ M, so that −1  R [K] = X – ∀R−1 [b], by the formula for switching between universal and direct relational image (4.5). This means that R−1 [K] is clopen, since by (a) it is equal to  Also, R is point-closed, since (4.9) shows the complement of the clopen set h(b). that, for any x ∈ X, R[x] is a closed down-set in Y . To finish the proof of (b), note that upward order-compatibility follows easily from the definitions, or also by an application of Exercise 4.3.2.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.3 Unary Operators

117

We now prove (c). Indeed, we will prove the following stronger fact, namely, that for any upward Priestley compatible relations R, S ⊆ X × Y , we have b] ⊆ ∀S −1 [ b]. S ⊆ R ⇐⇒ for every b ∈ M, ∀R−1 [

(4.12)

Note that the uniqueness (c) follows from (4.12), for if S is any upward Priestley  = ∀R−1 [ b] = h(b) b] for every b ∈ M, compatible relation satisfying (4.8), then ∀S −1 [ so S = R. The left-to-right direction of (4.12) is immediate from the definition of the universal image. For the converse, we will reason by contraposition and use the fact that R is point-closed. Suppose that S  R; pick x ∈ X and y ∈ Y such that y ∈ S[x] and y  R[x]. Since R[x] is closed, and also a down-set by order-compatibility, there exists b ∈ M such that R[x] ⊆  b and y   b, as follows from Proposition b], but 3.15. It now follows from the definition of universal image that x ∈ ∀R−1 [    x  ∀S −1 [b] since xSy but y  b. This concludes the proof of (4.12). Definition 4.40 The relation R defined in Proposition 4.39 is called the dual relation of the finite-meet-preserving function h. In Exercise 4.3.3 you are asked to use the uniqueness part of Proposition 4.39 to derive that, if X and Y are Priestley spaces, then every upward Priestley compatible relation R ⊆ X × Y is the dual relation of a unique finite-meet-preserving function. Proposition 4.39 is the crucial new ingredient for the following extension of the Priestley duality theorem (Theorem 3.31) to a larger collection of morphisms, namely all finite-meet-preserving functions. Theorem 4.41 The category of distributive lattices with finite-meet-preserving functions is dually equivalent to the category of Priestley spaces with upward Priestley compatible relations. The proof uses Proposition 4.39, combined with some techniques from category theory, and will be given in Section 5.4, p. 191. To finish this section, we draw a few further corollaries from Proposition 4.39. First, we show how it specializes to the Boolean case. Definition 4.42 Let X and Y be Boolean spaces. A relation R ⊆ X × Y is called Boolean compatible if it is point-closed and continuous, that is, for any clopen K ⊆ Y , the set R−1 [K] is clopen. Since the Priestley order on a Boolean space is trivial, note that a relation between Boolean spaces is upward Priestley compatible if, and only if, it is Boolean compatible according to Definition 4.42. Combining this observation with Propositions 4.39 and 3.36 allows us to deduce the following.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

118

4 Duality Methods

Corollary 4.43 Let A and B be Boolean algebras with dual spaces X and Y , respectively. Finite-meet-preserving functions B → A are in a one-to-one correspondence with Boolean compatible relations R ⊆ X × Y . For easy reference and future use, we also record the order-duals of Proposition 4.39 and Theorem 5.41, and the accompanying definitions. Definition 4.44 Let X and Y be Priestley spaces and let R ⊆ X × Y be a relation. We say that R is: • downward order-compatible if ≤ · R · ≤ ⊆ R, that is, for any x, x  ∈ X and y, y  ∈ Y , whenever x  ≤ x Ry ≤ y , we have x  Ry ; • lower Priestley continuous if, for every clopen down-set K ⊆ Y , the set R−1 [K] is clopen; and • downward Priestley compatible if R is downward order-compatible, lower Priestley continuous, and point-closed. Let X and Y be Priestley spaces and let R ⊆ X × Y be a downward Priestley compatible relation. We define the following function: ∃ R−1 : ClpD(Y ) → ClpD(X ) b → R−1 [b],

(4.13)

which we note is the same as X – ∀R−1 [Y – b] by (4.5). Conversely, any finitejoin-preserving function h : ClpD(Y ) → ClpD(X ) is equal to ∃R−1 for a unique downward Priestley compatible relation Rh ⊆ X × Y , which can be defined explicitly by Rh := {(x, y) ∈ X ×Y | for every b ∈ ClpD(Y ), if y ∈ b, then x ∈ h(b)}. (4.14) Proposition 4.45 Let L and M be distributive lattices with Priestley dual spaces X and Y , respectively. The assignments R → ∃ R−1 (4.13) and h → Rh (4.14) form a bijection between finite-join-preserving functions from M to L and downward Priestley compatible relations from X to Y . Proposition 4.45 has essentially the same proof as Proposition 4.39. Instead of redoing the entire proof, one may also appeal to order-duality to deduce this proposition from Proposition 4.39, see Theorem 5.41 in Chapter 5 and the remarks following it. We finish by examining two more special cases that may help elucidate the connection between the relations dual to finite-join- and finite-meet-preserving functions; in both of these cases, we examine a special setting where the two notions of dual relation interact with each other.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.3 Unary Operators

119

First, if f : L  M : g is an adjoint pair between distributive lattices, then the left adjoint f is finite-join-preserving and the right adjoint g is finite-meetpreserving (see Exercise 1.1.8). Denote by X and Y the Priestley dual spaces of L and M, respectively. By the results of this section, f has a dual downward Priestley compatible relation R f ⊆ Y × X and g has a dual upward Priestley compatible relation Rg ⊆ X × Y . The two relations are closely related: R f is the relational converse of Rg , see Exercise 4.3.5. Second, if h : M → L is a homomorphism between the distributive lattices L and M, then h has both an upward Priestley compatible dual relation Rh , because it preserves finite meets, and a downward Priestley compatible dual relation Sh , because it preserves finite joins. The intersection of Rh and Sh can be seen to be a functional relation, that is, for every x ∈ X, there is a unique y ∈ Y such that (x, y) ∈ Rh ∩ Sh . This relation Rh ∩ Sh is in fact equal to (the graph of) the continuous order-preserving function f : X → Y dual to h, which was defined in Section 3.2; see Exercise 4.3.6.

Exercises for Section 4.3 Exercise 4.3.1 Let h : M → L be a meet-preserving function between finite distributive lattices and let R be the relation defined in (4.6). (a) (b) (c)

Prove that R is upward order-compatible. Prove equation (4.7). Show directly (i.e., without referring to Proposition 4.39) that the assignment h → R is a bijection between meet-preserving functions from M to L and upward order-compatible relations from J (L) to J (M).

Exercise 4.3.2 Let X and Y be posets and R ⊆ X × Y a relation. Prove that R is upward order-compatible if, and only if, R−1 is downward order-compatible if, and only if, for any subsets S ⊆ X and T ⊆ Y , R[S] is a down-set and R−1 [T] is an up-set. Exercise 4.3.3 This exercise shows in more detail how the proof of Proposition 4.39 generalizes the proof sketch for the finite case given in the beginning of the section, and how one could naturally arrive at the compatibility conditions of Definition 4.38. It also establishes a link between duality for finite-meet-preserving maps and the canonical extension of such maps. Let X and Y be posets and let R ⊆ X × Y be a relation. (a)

Prove that the pair of functions R[−] : P (X )  P (Y ) : ∀R−1 [−] is an adjoint pair.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

120

(b)

4 Duality Methods

Using Exercise 4.3.2, show that it then follows that this adjunction restricts to a well-defined adjunction R[−] : D(X )  D(Y ) : ∀R−1 [−] if, and only if, R is upward order-compatible.

Further, assume that X and Y are Priestley spaces dual to distributive lattices L and M, respectively. (c) Now assume that R is upward order-compatible. Prove that R is upper Priestley continuous if, and only if, the function ∀ R−1 [−] : D(Y ) → D(X ) factors through the embeddings L → D(X ) and M → D(Y ), that is, if there exists a function  = ∀R−1 [ b] for all b ∈ M. Also show that such a h : M → L such that h(b) function h, if it exists, must preserve finite meets. (d) Assume that R is upward order-compatible and upper Priestley continuous, and let h : M → L be the function that exists according to item (c). Prove that R is point-closed if, and only if, R is the dual relation of h. (e) Show that, even if R is upward Priestley compatible, the function R[−] does not necessarily factor through L → D(X ) and M → D(Y ). (f) Conclude that, for any finite-meet-preserving function h : M → L, the function ∀ R−1 : D(Y ) → D (X ) is a completely meet-preserving function which extends  for all b ∈ M. b] = h(b) h, that is, ∀R−1 [ Note. The map ∀R−1 is called the canonical extension of the function h : M → L. For more information, see, for example, Gehrke and Jónsson (1994). It is possible to construct the canonical extension of a finite-meet-preserving function by algebraic means. This can then be used to give an alternative proof of Theorem 4.41. (g) Explain why the proof for the finite case, outlined earlier in this section, is a special case of the previous item. Exercise 4.3.4 This exercise shows that the correspondence between upward Priestley compatible relations and finite-meet-preserving functions generalizes Priestley duality for homomorphisms given in Chapter 3. Let X and Y be dual Priestley spaces of distributive lattices L and M, respectively. (a)

(b)

Prove that, if f : X → Y is a continuous order-preserving function, then R f := {(x, y) ∈ X × Y | f (x) ≥ y} is an upward Priestley compatible relation, which moreover has the property that R[x] has a maximum for every x ∈ X. Prove that, if R ⊆ X × Y is an upward Priestley compatible relation and R[x] has a maximum for every x ∈ X, then f R : X → Y defined by f R (x) := max R[x] is a continuous order-preserving function.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.4 Modal Algebras and Kripke Completeness

(c)

121

Prove that a finite-meet-preserving function f : M → L is a homomorphism if, and only if, the dual relation R is such that R[x] has a maximum for every x ∈ X.

Exercise 4.3.5 This exercise shows that “the dual relations of adjoint pairs are converse to each other.” Let f : L  M : g be an adjoint pair between distributive lattices and let X and Y be the Priestley dual spaces of L and M, respectively. Since f is a left adjoint, it preserves finite joins; let R f ⊆ Y × X denote the downward Priestley compatible relation dual to f . Similarly, let Rg ⊆ X × Y denote the upward Priestley compatible relation dual to g. Prove that (x, y) ∈ Rg if, and only if, (y, x) ∈ R f . That is, R f and Rg are converse relations. Hint. The first items of Exercise 4.3.3 can also be useful here. Exercise 4.3.6 Let h : M → L be a homomorphism between distributive lattices and let X and Y be the Priestley dual spaces of L and M, respectively. Denote by Rh ⊆ X × Y the upward Priestley compatible relation dual to h, viewed as a finite-meet-preserving function, and by Sh ⊆ X × Y the downward Priestley compatible relation dual to h, viewed as a finite-join-preserving function. Prove that the intersection Rh ∩ Sh of the two relations is a functional relation, and that this is the continuous order-preserving function dual to h, as defined in Section 3.2. 4.4 Modal Algebras and Kripke Completeness In this section, we show how the duality results of the previous section relate to Kripke’s possible world semantics for classical modal logic. We will see in Theorem 4.53 that the dual space of the free modal algebra gives a canonical model for the basic normal modal logic K. In this view, Kripke semantics for modal logic is obtained from the Stone–Jónsson–Tarski duality for operators that we developed in Section 4.3 by “forgetting the topology.” We begin by giving a minimal introduction to modal logic, limiting ourselves to the parts that are needed for understanding the connection to duality theory; for much more material on modal logic and duality we refer to classic textbooks in the field, such as Blackburn et al. (2001) and Chagrov and Zakharyaschev (1997). The basic normal modal logic, K, is an extension of classical propositional logic by a necessity operator, . Formally, modal formulas are terms built from propositional variables and the constants  and ⊥, using the binary operations ∨ and ∧, and unary operations ¬ and ; we use the common notational convention that unary operators bind more strongly than binary ones, that is, the notation p → ¬q denotes the formula (p) → (¬q), which is different from the formula (p → ¬q). We also consider the possibility operator  as an abbreviation, defined by de Morgan

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

122

4 Duality Methods

duality as ϕ := ¬¬ϕ, and implication may be defined by ϕ → ψ := ¬ϕ ∨ ψ. Note that, if one considers modal operators in the setting of positive logic (i.e., distributive lattices or Heyting algebras) rather than of classical logic (i.e., Boolean algebras), then the interplay between , , and the lattice operations is more subtle, and they are no longer interdefinable by formulas. In this section, for simplicity we only treat the classical, Boolean algebra-based case, although it is possible to develop dualities for modal algebras based on distributive lattices and Heyting algebras (Goldblatt, 1989; Celani and Jansana, 1999; Bezhanishvili, 1999). We will consider the distributive case in Example 5.42 and Exercise 5.4.3 in Chapter 5, and we also revisit duality for operators on distributive lattices in Section 6.4, where it will be related to a function space construction; see in particular Definition 6.21. A notion of derivability between modal formulas, K , may be defined by extending a Hilbert-style proof calculus for classical logic with one additional axiom, (p → q) → (p → q), and one additional rule: from K ϕ, infer K ϕ. We will not need to enter into details of the proof calculus here; see, for example, Chagrov and Zakharyaschev (1997, Section 3.6) for more details. Crucially for us is the characteristic property of K that, for any formulas ϕ and ψ, K (ϕ ∧ ψ) ↔ (ϕ ∧ ψ), and K  ↔ . In other words, the operation ϕ → ϕ yields a finitemeet-preserving function on the set of K -equivalence classes of modal formulas. This leads to the following definition. Definition 4.46 A modal algebra is a pair (B, ), where B is a Boolean algebra, and  : B → B is a finite-meet-preserving endofunction on B. A homomorphism from a modal algebra (B, B ) to a modal algebra ( A,  A ) is a homomorphism h : B → A such that, for every b ∈ B, h(B b) =  A h(b). Note that, if (B, ) is a modal algebra and V is a set of variables, and f : V → B is any function, then any modal formula ϕ with propositional variables drawn from V has a uniquely defined interpretation f˜(ϕ) in B, which may be defined inductively by setting f˜(v) := f (v) for v ∈ V , f˜() := , f˜(ϕ) :=  f˜(ϕ), f˜(ϕ ∨ ψ) := f˜(ϕ) ∨ f˜(ψ), f˜(¬ϕ) := ¬ f˜(ϕ), etc. In algebraic terms, f˜ is the unique extension of f to a homomorphism from the term algebra over V to B. It may be proved that the set of K-derivable equivalence classes of modal formulas in a fixed set of variables V is, up to isomorphism, the free modal algebra over V . The proof is a straightforward adaptation to the modal case of the construction of the Lindenbaum–Tarski algebra for propositional logic, which we saw when we constructed the free distributive lattice in Section 4.1, see Remark 4.5; for all the details in the modal setting, see, for example, Chagrov and Zakharyaschev (1997, Section 7.5) or Blackburn et al. (2001, Section 5.2). From this fact, one obtains the following theorem, which provides the foundation for algebraic modal logic, where it is known as the “algebraic completeness theorem for K.”

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.4 Modal Algebras and Kripke Completeness

123

Theorem 4.47 Let ϕ and ψ be modal formulas with propositional variables among x 1, . . . , x n . Then, K ϕ ↔ ψ if, and only if, for every function f : {x 1, . . . , x n } → B, with B a modal algebra, we have f˜(ϕ) = f˜(ψ). In particular, K ϕ if, and only if, f˜(ϕ) =  for every interpretation f . Proof See, for example, Chagrov and Zakharyaschev (1997, Theorems 7.43 and 7.44) or Blackburn et al. (2001, Theorem 5.27).  We now show how this algebraic completeness theorem for K may be combined with the duality in Theorem 4.41 of Section 4.3 to obtain easy proofs of Kripke completeness theorems for K. We first give the definition of Kripke semantics. Definition 4.48 A (discrete) Kripke frame is a pair (X, R), with X a set and R ⊆ X × X a relation. A Kripke Boolean space is a triple (X, τ, R), with (X, τ) a Boolean space and R ⊆ X × X a compatible relation on X. For a set of propositional variables V , a valuation to a Kripke frame (X, R) is a function c : X → 2V . When (X, τ, R) is a Kripke Boolean space, a valuation c : X → 2V is called admissible if, for every v ∈ V , the set {x ∈ X | c(x)(v) = 1} is clopen. A (discrete) Kripke model is a tuple (X, R, c) where c is a valuation, and a Kripke Boolean model is a tuple (X, τ, R, c) where (X, τ, R) is a Kripke Boolean space and c is an admissible valuation. Kripke models can be used to define a concrete, set-based semantics for modal logic, generalizing truth tables for propositional logic, as follows. Fix a set of variables V . For any Kripke model (X, R, c), a forcing or truth relation, |= is defined between points of X and modal formulas with variables in V , as follows. For any x ∈ X and v ∈ V , define x |= v if, and only if, c(x)(v) = 1, and also define x |=  to always hold. Then, extend |= inductively by defining, for any modal formulas ϕ, ψ, that x |= ϕ ∨ ψ if, and only if, x |= ϕ or x |= ψ; x |= ¬ϕ if, and only if, it is not the case that x |= ϕ; and finally, def

x |= ϕ ⇐⇒ for every y ∈ X, if x Ry, then y |= ϕ.

(4.15)

If x |= ϕ, then we say that ϕ holds or is true at x ∈ X. This definition in particular applies to a Kripke Boolean model (X, τ, R, c), where the further condition that the valuation c must be admissible implies that the truth set of ϕ, ϕ := {x ∈ X | x |= ϕ}, is clopen, for every modal formula ϕ; also see Proposition 4.52 and Exercise 4.4.1. If ϕ is a modal formula with variables in V and (X, R, τ, c) is a Kripke Boolean

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

124

4 Duality Methods

model, then ϕ is valid on the model provided that ϕ is true at every x ∈ X. The formula ϕ is valid on a Kripke Boolean space if it is valid for every admissible valuation. On the other hand, ϕ is valid on a (discrete) Kripke frame (X, R) it is valid under every valuation. Clearly, if (X, R, τ) is a Kripke Boolean space and ϕ is a formula that is valid on the underlying discrete Kripke frame (X, R), then ϕ is in particular valid on the Kripke Boolean space (X, R, τ). The converse is not true in general; for an example, see Exercise 4.4.4. Example 4.49 Consider the Kripke frame (N, h(ai ), contradicting the definition of h(ai ). The “in particular” statement now follows immediately from the fact that U (X ) − is the Boolean subalgebra of P (X ) generated by U (X ) (Corollary 3.35).  We now show a somewhat more technically involved lemma, which in particular establishes the inclusion A ⊆ L − , and thus the equality of Boolean algebras L − = A claimed in Theorem 4.79. Lemma 4.81 Suppose that a ∈ P (X ) has alternation height h ∈ N. Then, there exist a0, . . . , ah and b0, . . . , bh in U (X ) such that a=

h 

(an – bn ).

n=0

Proof Throughout this proof, we slightly extend the definition of alternating chain, and we will call a sequence x 1 ≤ y1 ≤ · · · ≤ x n ≤ yn ≤ x n+1 of elements of X with x i ∈ a for 1 ≤ i ≤ n + 1 and yi  a for 1 ≤ i ≤ n an alternating chain for a of height n + 12 .

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.7 Boolean Closure and Alternating Chains

151

We inductively define the following sequence of elements of U (X ): a0 := ↑a;

b0 := ↑(a0 ∩ a c );

an+1 := ↑(bn ∩ a); and

bn+1 := ↑(an+1 ∩ a c ) for 0 ≤ n ≤ h.

We now first show by induction on n that, for any n ≥ 0 and any x ∈ X, • x ∈ an if, and only if, there exists an alternating chain for a of height n + 12 whose last point is ≤ x, and • x ∈ bn if, and only if, there exists an alternating chain for a of height n + 1 whose last point is ≤ x. For n = 0, note that an alternating chain of height 12 is just a single element of a, which gives the first item. For the second item, unraveling the definition of b0 , we see that x ∈ b0 if, and only if, there exist y0  a and x 0 ∈ a with x 0 ≤ y0 ≤ x, as required. Now, let n ≥ 1 be arbitrary. For any x ∈ X, by definition we have x ∈ an if, and only if, there exists x  ∈ bn−1 ∩ a with x  ≤ x. Suppose that x ∈ an . By induction, since x  ∈ bn−1 , there is then an alternating chain for a of height n whose last point is ≤ x . Since x  ∈ a, extending this sequence by adding x  to the end gives an alternating chain for a of height n + 12 and x  ≤ x. Conversely, if x 1 ≤ y1 ≤ · · · ≤ x n ≤ yn ≤ x n+1 is an alternating chain for a of height n + 12 with x n+1 ≤ x, then x n+1 ∈ a, and the alternating chain x 1 ≤ y1 ≤ · · · ≤ x n ≤ yn shows, by induction, that x n+1 ∈ bn−1 , so that x ∈ ↑(bn−1 ∩ a) = an , as required. The proof of the second item is very similar. Let us now write c0 := ∅ and, for each 1 ≤ n ≤ h + 1, cn :=

n−1 

(ai – bi ).

i=0

The statement of the lemma is that a = ch+1 . Toward proving this, we will show that, for each 0 ≤ n ≤ h + 1, cn ⊆ a ⊆ cn ∪ an .

(4.29)

Note that the fact that the alternation height of a is h implies, using the characterization above, that bh = ah+1 = ∅. Thus, (4.29) for k = h + 1 will give the desired equality a = ch+1 . For the first inclusion of (4.29), it suffices to show that an – bn ⊆ a for every 0 ≤ n ≤ h. Indeed, if x ∈ an – bn , pick an alternating chain of height n + 12 for a whose last point is ≤ x. Since x  bn , adding x to the end of this alternating chain does not give an alternating chain of height n + 1, so we must have x ∈ a. For the second inclusion of (4.29), we proceed by induction on n. When n = 0, simply note that indeed a ⊆ a0 . For the inductive step, suppose that x ∈ a. We distinguish two cases, according to whether or not x ∈ bn−1 .

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

152

4 Duality Methods

Case 1: x ∈ bn−1 . Then, there exists an alternating chain for a of height n whose last point is ≤ x. Since x ∈ a, we can extend this sequence to an alternating chain of height n + 12 whose last point is x, so x ∈ an . Case 2: x  bn−1 . Then, if x ∈ an−1 , we are done because an−1 – bn−1 ⊆ cn by definition. Assume x  an−1 . By the inductive hypothesis, we have a ⊆ cn−1 ∪ an−1 ,  so x ∈ cn−1 . Clearly, cn−1 ⊆ cn , so we conclude x ∈ cn . Exercises for Section 4.7 Exercise 4.7.1 This exercise is about the collection of elements {ai, bi }i ∈N that we construct for an element a ∈ P (X ) in the proof of Lemma 4.81. (a) (b)

Show that a0 ⊇ b0 ⊇ a1 ⊇ · · · ⊇ bn for all n ∈ N. Show that n 

(ai – bi ) = a0 – (b0 – (a1 – · · · (an – bn ) . . . )).

i=0

(c)

For an element a ∈ P (X ), we call a0 ⊇ b0 ⊇ a1 ⊇ · · · ⊇ bn such that ai, bi ∈ U (X ) and h  (an – bn ) a= n=0

a difference chain for a and we order difference chains for a by coordinatewise inclusion (if one is shorter than the other, then we consider it extended with empty sets). Show that if a has a difference chain then it has a least such and it is the one we define in the proof of Lemma 4.81. Exercise 4.7.2 Suppose H is a co-Heyting algebra, and X its dual space. Consider X with the opposite of the Priestley order so that H is isomorphic to the lattice of clopen up-sets via a →  a. Let V ⊆ X be a clopen in X. Show that the difference chain for V as defined in Exercise 4.7.1 consists entirely of clopen up-sets of X. Conclude that each element of H − has a least difference chain over H. Exercise 4.7.3 Show that for an arbitrary bounded distributive lattice L, each element a ∈ L − is given by a difference chain a0 > b0 > · · · an > bn of elements of L such that a = a0 – (b0 –(a1 – · · · (an – bn ) . . . )) =

h 

(an – bn ),

n=0

and that this gives rise to a well-defined notion of alternation height for elements a ∈ L − over L. Hint. If a ∈ L − , then a ∈ F − , where F is some finite sublattice of

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

4.7 Boolean Closure and Alternating Chains

153

L. The finite sublattice F is then in particular a co-Heyting algebra, and provides a least difference chain for a over F, which is also a difference chain for a over L. Exercise 4.7.4 Give an example of a distributive lattice L, and an element a ∈ L − to show that there is not a least difference chain for a in general. Exercise 4.7.5 Show that the closed down-sets of the dual Priestley space of a distributive lattice L form a co-Heyting algebra H containing L, and that the alternation height a ∈ L − is the same over either L or H. Notes for Chapter 4 The theory of canonical extensions allows one to make more precise the idea, mentioned above Proposition 4.39, that the duality for operators treated in this chapter is an extension of the finite case; see also Exercise 4.3.1. In that theory, any distributive lattice L embeds into a “finite-like” lattice L δ and any finite-meetpreserving function h : M → L is shown to lift to a completely meet preserving function h δ : M δ → L δ , which then yields the dual relation R from the dual space of X to the dual space of Y , exactly by the ideas from the finite duality outlined in Section 4.3. This algebraic definition of the dual relation was at the core of the work in Jónsson and Tarski (1951, 1952) in the case of Boolean algebras. Gehrke and Jónsson (1994, 2004) generalize this theory from Boolean algebras to distributive lattices and in addition study generalizations of the theory to arbitrary operations that do not need to preserve joins or meets. A property of a lattice of operators is called canonical if it is preserved under canonical extension. The study of canonicity and its connections to first order definability is a major research theme in modal logic, see, for example, Goldblatt et al. (2003) or Hodkinson and Venema (2005). The field of Sahlqvist theory provides syntactic conditions on properties of lattices with operators that guarantee canonicity, see, for example, Blackburn et al. (2001, Chapter 5), Gehrke et al. (2005), Conradie et al. (2014), and Conradie and Palmigiano (2020). The structures that we call “Kripke Boolean spaces” in Section 4.4 are typically presented in the modal logic literature in the form of descriptive general frames, which give an isomorphic category. The latter however avoids having to speak explicitly about topology, see, for example, Blackburn et al. (2001, Chapter 5) or Chagrov and Zakharyaschev (1997, Chapter 8). We prefer to make the Boolean topology on general frames explicit, because we think it clarifies the mathematical content of the descriptive general frame semantics, and in particular the link with Stone duality.

https://doi.org/10.1017/9781009349680.005 Published online by Cambridge University Press

5 Categorical Duality

In this chapter, we introduce notions from category theory and use them to rephrase the duality theorems that we have seen so far. For our purposes, category theory provides a general language that allows us to talk about duality from a higher level viewpoint. We limit ourselves to explaining only those notions from category theory that we need in this book; this is by no means a complete introduction to category theory. In particular, we do not go into deep connections between the topological duality theory developed in this book and the rich category theory of frames and toposes; for more on this, see, for example, Johnstone (1986), Makkai (1993), Mac Lane and Moerdijk (1992), Hofmann et al. (2014), and Caramello (2017). Since we want to keep category-theoretic preliminaries to a minimum, all the category theory treated in this chapter is entirely standard and well known, and can be found in a lot more detail in, for example, Mac Lane (1971), Awodey (2010), and Leinster (2014).

5.1 Definitions and Examples of Categories A category formalizes a class of mathematical objects, the morphisms between them, and how the morphisms compose. For example, the class of sets and functions, with usual function composition, is a category, but so is the class of sets and binary relations, with relational composition. A category is a two-sorted structure: it has an objects sort and a morphisms sort. Every morphism f in a category has an associated domain object, dom( f ), and codomain object, cod( f ). For every “composable” pair of morphisms ( f , g), that is, a pair such that dom(g) = cod( f ), there is a morphism, g ◦ f , the composition of g after f , and this operation of composition is required to be associative. Finally, every object A in a category has an associated identity morphism 1 A, which acts as a neutral element for the composition. We now give the formal definition. Although

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.1 Definitions and Examples of Categories

155

this definition may look lengthy and abstract at first, the conditions are natural and easily seen to be satisfied in the examples that follow. Definition 5.1 A category C is a tuple (ob C, mor C, dom, cod, ◦, 1(−) ), where ob C and mor C are classes of objects and morphisms, dom and cod are assignments from mor C to ob C, 1(−) is an assignment from ob C to mor C, and ◦ is an assignment from {(g, f ) | dom(g) = cod( f )} to mor C, with the following properties: (a)

For any f , g ∈ mor C with dom(g) = cod( f ), dom(g ◦ f ) = dom( f ) and cod(g ◦ f ) = cod(g).

(b)

For any f , g, h ∈ mor C, if dom(g) = cod( f ) and dom(h) = cod(g), then h ◦ (g ◦ f ) = (h ◦ g) ◦ f .

(c)

For any A ∈ ob C, dom(1 A ) = A = cod(1 A ).

(d)

For any f ∈ mor C with dom( f ) = A and cod( f ) = B, f ◦ 1 A = f and 1B ◦ f = f .

While Definition 5.1 gives the official definition of the structure usually called a category, alternative definitions are possible. In particular, the definition of category does not need to be two sorted, as the structure of a category is entirely determined by its class of morphisms: The objects can be encoded in the class of morphisms using the identity morphisms; see, for example, Mac Lane (1971, p. 9) or Adámek et al. (1990, Definition 3.53, p. 42). The structure of categories is rich, due to many derived notions, which we begin to introduce now. Let C be a category. For any A, B ∈ ob C, the Hom-class from A to B in C is the collection of morphisms f with dom( f ) = A and cod( f ) = B; this class is denoted by HomC ( A, B), or simply C( A, B). We also use the notation f : A → B to mean f ∈ HomC ( A, B) when the ambient category is clear. The notation ◦ and parentheses are often omitted: For example, we write h g f for the morphism h◦ (g ◦ f ) = (h◦g) ◦ f . We also sometimes denote the identity morphism on an object A by id A instead of 1 A. Note that the components defining a category are not assumed to be sets. This level of generality is necessary, because, for example, the collection of all sets is not a set itself, and we want to be able to consider the category of sets. Unless explicitly mentioned otherwise, all the categories we work with in this book are locally small, which means that, for any fixed pair of objects A, B in the category, the Hom-class from A to B is a set. A category is called small if the class of morphisms (and, hence, also the class of objects) is a set, and large otherwise.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

156

5 Categorical Duality

Example 5.2 The following are examples of categories that we have already seen. In each example, we only state what ob C and mor C are; composition and identity are what you expect. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j)

The category DL of distributive lattices and their homomorphisms. The category Top of topological spaces and continuous functions. The category Priestley of Priestley spaces and continuous order-preserving functions. The category BA of Boolean algebras and their homomorphisms. The category BoolSp of Boolean spaces and continuous functions between them. The category HA of Heyting algebras and their homomorphisms. The category Esakia of Esakia spaces and continuous p-morphisms. The categories DL f and BA f of finite distributive lattices and finite Boolean algebras, respectively, and homomorphisms between them. The category Set of sets and functions; the category Set f of finite sets and functions between them. The category Pos of posets and order-preserving functions; the category Pos f of finite posets and order-preserving functions between them.

The categories described in the above example correspond to the way we motivated categories above, as “mathematical universes.” In technical terms, these are large categories that are concrete, meaning intuitively that each object is a set (with additional structure) and each morphism is a (special kind of) function between the sets underlying the objects. We will be a bit more precise about this concept of a “concrete category” in Example 5.9(c). However, there are other, more abstract examples of categories. Example 5.3 We give two classes of typical examples of categories, thin categories and monoids. (a)

(b)

Let (P, ≤) be a preorder. There is a category P with P as the set of objects and ≤ as the set of morphisms; that is, a morphism is a pair (p, q) ∈ P × P with p ≤ q, dom((p, q)) := p, and cod((p, q)) := q. Exercise 5.1.1 asks you to define the additional structure that makes P into a category. Categories P that arise in this way are called thin; they are characterized by the property that there is at most one morphism between any pair of objects. Let (M, ·, 1) be a monoid, that is, a set with an associative operation · such that 1 · m = m = m · 1 for all m ∈ M. There is a category M with a single object ∗ and M as the set of morphisms. The domain and codomain of any morphism is by definition ∗, the identity morphism 1∗ on the object ∗ is 1,

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.1 Definitions and Examples of Categories

157

and composition is defined by ·. The axioms of a monoid immediately give that M is indeed a category. Both of these examples of categories are in a sense degenerate, but in two different directions: few morphisms (preorders) and one object (monoids). A general category can thus be thought of as either “a multi-arrow preorder” or “a multi-object monoid.” One more central notion in categories, that we introduce now because we will need it throughout this chapter, is that of an isomorphism. Definition 5.4 Let f : A → B be a morphism in a category C. A two-sided inverse to f is a morphism g : B → A in C such that g ◦ f = 1 A and f ◦ g = 1B . The morphism f is called an isomorphism if a two-sided inverse exists for it. Two objects A and B are called isomorphic if there exists an isomorphism from A to B. If f is an isomorphism, then the inverse of f is unique, and is denoted by f −1 . This morphism f −1 is also an isomorphism, and the notion of being isomorphic is in fact an equivalence relation on the objects of C (see Exercise 5.1.4). 1 A skeleton of a category C is a full subcategory S of C such that each object is isomorphic to exactly one object of S (see Exercise 5.2.3). Example 5.5 In any of the concrete categories from Example 5.2, “isomorphism” as defined in Definition 5.4 means exactly what one would expect. In particular, an isomorphism between sets is just a bijection and an isomorphism between Priestley spaces is an order homeomorphism. In a thin category P as in Example 5.3(a), two objects are isomorphic if, and only if, they are equivalent in the sense of Exercise 1.1.5. In a single-object category M as in Example 5.3(b), an isomorphism corresponds to an invertible element of the monoid. In particular, the monoid M is a group if, and only if, every morphism of M is an isomorphism. More generally, a category G is called a groupoid if every morphism in G is an isomorphism.

Exercises for Section 5.1 Exercise 5.1.1 Let (P, ≤) be a preordered set. Define ob P := P and mor P := ≤. For any (p, q) ∈ mor P, let dom((p, q)) := p and cod((p, q)) := q. (a) 1

Why does there exist, for every p ∈ P, a unique morphism 1 p with domain and codomain p?

Note that, since ob C is not necessarily a set, while “equivalence relations” are usually taken to be subsets, when we say “equivalence relation” in this chapter, more formally we mean: a binary predicate that satisfies reflexivity, symmetry, and transitivity; but we will not need to worry about size issues.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

158

(b) (c) (d)

5 Categorical Duality

For any (p, q), (q, r) ∈ mor P, define (q, r) ◦ (p, q) := (p, r). Why is (p, r) a morphism in P? Verify that the structure defined above turns P into a category in which there is at most one morphism between any two objects. Conversely, if C is a small category in which there is at most one morphism between any two objects, define a preorder ≤ on the set ob C such that C is the category associated to the preorder (ob C, ≤).

Exercise 5.1.2 Let C be a small category with one object. Define a monoid M such that C is the category associated to the monoid M, as in Example 5.3(b). Exercise 5.1.3 For any sets X, Y, Z and relations R ⊆ X × Y and S ⊆ Y × Z, recall that the relational composition of R and S is the relation R · S ⊆ X × Z defined by R · S := {(x, z) ∈ X × Z | ∃ y ∈ Y such that (x, y) ∈ R and (y, z) ∈ S}. Verify that the following properly defines a category Rel: let ob Rel be the class of sets, and for any sets X, Y , let HomRel (X, Y ) be the set of relations R ⊆ X × Y . For any set X, let 1X be the diagonal relation on X. Finally, for R ∈ HomRel (X, Y ) and S ∈ HomRel (Y, Z ), define S ◦ R := R · S. Exercise 5.1.4 Let f : A → B be a morphism in a category C. (a) (b) (c)

Prove that if g and g  are two-sided inverses to f , then g = g . Suppose that f is an isomorphism. Prove that its two-sided inverse f −1 is also an isomorphism. Define a relation ∼ on ob C by: A ∼ B if, and only if, A is isomorphic to B. Prove that ∼ is an equivalence relation. 5.2 Constructions on Categories

Categories can be studied from two different perspectives, both of which are important; one is external, the other internal. From the external perspective, one regards categories themselves as mathematical objects. From this point of view, one may consider mappings between categories and constructions on categories. For example, in this section we will describe the opposite category, subcategories of a category, functors between categories, adjunctions, isomorphisms, and equivalences between categories and, most importantly for our purposes, dualities. From the internal perspective, the structure of a category, viewed as a universe for performing mathematical operations, allows us to mimic many natural constructions that we know from mathematics within a given, fixed category. In the previous

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

159

section, we already saw in Definition 5.4 that the familiar notion of isomorphism between mathematical objects can be defined internally in any category. We will see in the next section, Section 5.3, that within many of the categories relevant to us, there is an analogue of “injective” and “surjective” map, of product, disjoint unions, power sets, and more. Many interesting results in category theory come from the interplay between the two perspectives. As a case in point, duality theory uses an external construction (a duality) to obtain knowledge about internal constructions (for example, quotients and subobjects) of the categories at hand. As an aside, we note that many of the constructions that we describe here as external constructions on categories can also be viewed as internal constructions within a category, namely, the category Cat of categories (to be properly defined shortly). This phenomenon, which is both beautiful and a bit jarring, will mostly be ignored in this chapter. We recall that our aim in this chapter is not to give the reader a complete course in category theory, but to only introduce categories to the extent that a working duality theorist needs them.

The Opposite Category For a category C, the opposite category Cop is the category with the same objects, morphisms, and identity as the category C, but the domain and codomain assignments are interchanged, and the order of writing composition is reversed. More formally: Definition 5.6 Let C = (ob C, mor C, dom, cod, ◦, 1− ) be a category. The opposite category Cop is defined as (ob C, mor C, cod, dom, •, 1− ), where, if f : A → B, g : B → C are composable morphisms in C with composite g ◦ f , then the composition of the corresponding morphisms g : C → B and f : B → A in Cop , f • g, is defined to be the morphism C → A in Cop corresponding to g ◦ f . Thus, for any objects A, B, we have by definition HomCop ( A, B) = HomC (B, A).

Subcategories If C and D are categories with ob D ⊆ ob C and mor D ⊆ mor C and the assignments dom, cod, ◦, and 1(−) of D are restrictions of those of C, then D is called a subcategory of C. An important special case of a subcategory is that of a full subcategory. A subcategory D of C is called a full subcategory if HomD ( A, B) = HomC ( A, B) for every A, B ∈ ob D. If D is a collection of objects in a category C, then we can always consider the full subcategory of C whose objects are in D.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

160

5 Categorical Duality

For example, the categories Set f , DL f , and BA f are full subcategories of Set, DL, and BA, respectively. Also, BA is a full subcategory of DL, using the fact that every lattice homomorphism between Boolean algebras is a Boolean homomorphism (see Exercise 1.2.11(b)). The category HA is a non-full subcategory of DL: Not every lattice homomorphism between Heyting algebras is a Heyting homomorphism (see Exercise 4.6.5). A subcategory may only restrict morphisms, while keeping the same class of objects. For example, the category Rel introduced in Exercise 5.1.3 has Set as a non-full subcategory: when moving from Rel to Set, the objects stay the same, but the morphisms are those relations that are functions. The category Heyt f of finite Heyting algebras is an example of a non-full subcategory of DL f : Every finite distributive lattice is a finite Heyting algebra, so the objects are the same, but not every distributive lattice homomorphism preserves the Heyting implication (see Exercise 4.6.5). Functors and Natural Transformations The correct notion of “homomorphism” between categories is that of a functor. We have already seen several examples of functors. For example, in Definition 3.21, we gave a functor from DL to Priestleyop : We associated to any object L of the category DL an object X L of the category Priestley, and, in Proposition 3.27, to any morphism h : L → M of the category DL a morphism f : X M → X L of the category Priestley, that is, a morphism f : X L → X M in Priestleyop . Definition 5.7 Let C and D be categories. A functor F from C to D is a pair of assignments, ob C → ob D and mor C → mor D such that the following properties hold: (a) (b) (c)

for any f ∈ mor C, dom(F ( f )) = F (dom( f )) and cod(F ( f )) = F (cod( f )); for any A ∈ ob C, F (1 A ) = 1F A; and for any composable pair ( f , g) ∈ mor C, F (g ◦ f ) = F (g) ◦ F ( f ).

Note a slight abuse of notation in this definition: Both the object and the morphism assignment are denoted by F, even though they are strictly speaking two separate components of the functor F. This notation rarely leads to confusion. Let F : C → D be a functor. Then, F is called • full if, for any two objects A, B of C, the assignment f → F ( f ) is surjective as a map from HomC ( A, B) to HomD (F A, F B), • faithful if, for any two objects A, B of C, the assignment f → F ( f ) is injective as a map from HomC ( A, B) to HomD (F A, F B), and • essentially surjective if for every object B in D, there exists an object A ∈ C such that F A is isomorphic to B in D.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

161

Functors can be composed in the obvious way, and on any category C there is an obvious identity functor, denoted 1C . In this way, as mentioned above, the collection of locally small categories and functors between them is itself again a category, Cat. We will not pursue this point any further here. Example 5.8 We give some first examples of functors. (a)

(b) (c)

If P and Q are preorders, viewed as categories as in Example 5.3, then a functor f : P → Q is essentially the same thing as an order-preserving function: Given a function on the underlying sets, the only non-trivial requirement for it to extend to a functor between the associated categories P and Q is that, if p ≤ p, then there must be a morphism from f (p) to f (p ) in Q, in other words, f (p) ≤ f (p ). If M and N are monoids, viewed as categories as in Example 5.3, then a functor f : M → N is essentially the same thing as a monoid homomorphism. There is a functor U : DL → Set that sends any distributive lattice L to its underlying set, and any homomorphism h : L → M to itself. The functor U is clearly faithful, but not full. This functor U is called the forgetful functor, and in fact such a functor exists in a much wider setting, including any category of “algebraic structures.” While the functor may look rather trivial, it is important for defining free objects categorically, as we will see in Example 5.17(a). Categories that admit a well-behaved forgetful functor to Set are sometimes called concrete categories, and enjoy special properties, also see Adámek et al. (1990).

Example 5.9 We recall several examples of functors from a category C to a category Dop that we have already seen in Chapters 1, 3, and 4. Such functors are sometimes called contravariant functors from C to D, and in this context, an actual functor from C to D is called a covariant functor. (a)

(b)

(c)

The functor D : Pos f → (DL f ) op is defined by sending a finite poset to its lattice of down-sets, and an order-preserving function f : P → Q to the lattice homomorphism D( f ) : D(Q) → D(P), D → f −1 (D). The functor P : Set f → (BA f ) op is defined by sending a finite set to its Boolean power set algebra and a function f : X → Y to the Boolean homomorphism P ( f ) : P (Y ) → P (X ), S → f −1 (S). If we view the category Set f as the full subcategory of Pos f consisting of the finite posets with the discrete order, then P is the restriction of D to this full subcategory. The functor ClpD : Priestley → DLop is defined by sending a Priestley space X to the lattice ClpD(X ) of clopen down-sets of X and a continuous order-preserving function f : X → Y to the lattice homomorphism ClpD( f ) : ClpD(Y ) → ClpD(X ), D → f −1 (D). If we view Pos f as the full

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

162

(d)

(e)

5 Categorical Duality

subcategory of Priestley consisting of the finite Priestley spaces (see Exercise 5.2.1), then ClpD restricts to the functor D in item (a) above. The functor Clp : BoolSp → BAop is defined by sending a Boolean space X to the Boolean algebra Clp(X ) of clopen subsets of X and a continuous function f : X → Y to the Boolean homomorphism f −1 : Clp(Y ) → Clp(X ). If we view Set f as the full subcategory of BoolSp consisting of the finite discrete spaces, then Clp restricts to the functor P in item (b) above. Also, if we view BoolSp as a subcategory of Priestley by equipping a Boolean space with the trivial partial order, then the functor Clp is the restriction of ClpD of item (c) above. Write PriestleyR↑ for the category whose objects are Priestley spaces and whose morphisms are upward Priestley compatible relations; see Definition 4.38. Write DL∧ for the category whose objects are distributive lattices and whose morphisms are finite-meet-preserving functions. The results in Section op 4.3 show in particular that there is a functor ClpD : PriestleyR↑ → DL∧ , which acts on objects as the functor described under item (c) above, and on morphisms sends a relation R ⊆ X × Y to the finite-meet-preserving function ∀ R−1 .

In each of the above examples, there is also a functor in the other direction. For example, recall from Theorem 1.25 that there is a functor J : (DL f ) op → Pos f which is defined by sending a finite distributive lattice L to its poset of joinirreducibles J (L), and a homomorphism h : L → M to the restriction of its lower adjoint, which is an order-preserving function J (M) → J (L). How are the functors D and J related to each other? A first guess might be that they form an isomorphism between categories, in the sense that they are mutually inverse to each other, but this is not literally the case. (We give the notion of isomorphism of categories at the end of this section.) The composition of the functors D and J is “almost” the identity, but only up to an isomorphism between objects: For any lattice L, there is an isomorphism α L between L and D(J (L)). While the isomorphism α L has the lattice L as a parameter, its definition is “consistent” as L varies. To precisely define an equivalence of categories, we need to make precise what we mean by “consistent” in the previous sentence. To this end, we now introduce the notion of natural transformation. A natural transformation should be thought of as a morphism between two functors, and, as we shall see, the family of isomorphisms α := (α L ) L ∈ob DL f is an example. The family α is a special kind of natural transformation, because each of its components is an isomorphism; this is not required in the general definition of natural transformation. Definition 5.10 Let F and G be two functors from a category C to a category D. A natural transformation from F to G, notation ϕ : F ⇒ G, is an (ob C)-

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

163

indexed collection ϕ = (ϕ A ) A∈ob C of morphisms in D that satisfies the following properties: (a) (b)

for every A ∈ ob C, dom(ϕ A ) = F A and cod(ϕ A ) = G A; and for every morphism f : A → A in C, we have G f ◦ ϕ A = ϕ A ◦ F f ; that is, the following square commutes: FA

Ff

ϕ A

ϕA GA

F A

Gf

G A .

A natural isomorphism is a natural transformation all of whose components are isomorphisms in the category D. Under this definition, the family α = (α L ) L ∈DL f defined above is an example of a natural isomorphism α : 1DL f ⇒ D ◦ J , where we recall that 1DL f denotes the identity functor on the category DL f . Natural transformations are the morphisms in a (large) category of functors, that is, there is a composition of natural transformations and an identity natural transformation 1F : F ⇒ F, for any functor F. Moreover, when ϕ : F ⇒ G is a natural transformation and K, L are functors that can be post-composed and pre-composed with F, then we also have well-defined natural transformations K ϕ : K F ⇒ KG and ϕ L : F L ⇒ GL. In this category, a natural transformation is a natural isomorphism if, and only if, it is an isomorphism in the category, that is, if there exists a natural transformation that is its two-sided inverse. For more precise statements, see Exercise 5.2.4. Example 5.11 Let P and Q be preorders. Recall from Example 5.9 that functors from P to Q are (given by) order-preserving functions. Since there is at most one morphism between any two objects of P, or of Q, there is at most one natural transformation between any two functors in this setting. If f , g : P → Q are two order-preserving functions, viewed as functors, then there exists a natural transformation α : f → g if, and only if, f is pointwise below g, that is, f (p) ≤ g(p) for every p ∈ P. For monoids M and N, a natural transformation between two functors (i.e., homomorphisms) f , g : M → N is given by an element α ∈ N such that, for every m ∈ M, α f (m) = g(m)α. Equivalences and Dualities Finally, we come to the categorical notion that is central to this book.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

164

5 Categorical Duality

Definition 5.12 Let C and D be categories. A pair of functors F : C  D : G is called an equivalence between C and D if there exist natural isomorphisms α : 1C ⇒ GF and β : FG ⇒ 1D . An equivalence between C and Dop is called a dual equivalence or duality between C and D. The categories C and D are called equivalent if there exists an equivalence between them and C and D are called dually equivalent or simply dual if there exists a dual equivalence between them. In this definition, since the notion of natural isomorphism is symmetric, the direction of α and β clearly does not matter, and the asymmetric choice, where α has the identity functor as codomain and β has the identity functor as domain, may look a bit strange. However, we will see shortly that the notion of equivalence is a special case of the notion of adjunction, where the direction of the natural transformations does matter. As is common in the literature, we formulated Definition 5.12 to resemble the definition of adjunction (Definition 5.15) as much as possible. The following theorem, whose proof relies on the axiom of choice, is very useful when proving that a functor is part of an equivalence, as it allows us to avoid explicitly defining natural isomorphisms, but instead to just check three properties for one of the two equivalence functors. One may view this theorem as a categorified version of the fact that if a function f : C → D between sets is bijective, then it has a two-sided inverse. Theorem 5.13 Let F : C → D be a functor. The following are equivalent: (i) (ii)

There exists a functor G : D → C such that (F, G) is an equivalence between C and D. The functor F is full, faithful, and essentially surjective.

Proof We leave the implication (i) ⇒ (ii) to the reader (see Exercise 5.2.5). Conversely, suppose that F is full, faithful, and essentially surjective. Since F is essentially surjective, for every object B in D, pick an object GB in C and an isomorphism β B : F (GB) → B. Let g : B → B  be a morphism in D and define g  : F (GB) → F (GB  ) to be the composite map ( β B ) −1 ◦ g ◦ β B . Since F is full and faithful, there exists a unique morphism, Gg : GB → GB , such that F (Gg) = g  = ( β B ) −1 ◦ g ◦ β B . In other words, applying β B to both sides of this equality, Gg is the unique morphism GB → GB  such that β B F (Gg) = g β B . (We will often omit the symbol ◦ in the remainder of this proof to improve readability.) We show that the assignment G : D → C given by G(B) = GB and G(g) = Gg is a functor. For any object B in D, notice that β B F (1GB ) = β B 1F (GB) = β B = 1B β B . Thus, by the uniqueness in the definition of G on morphisms, G(1B ) = 1GB . Now, let g1 : B1 → B2 and g2 : B2 → B3 be a pair of composable morphisms in D. We will prove that G(g2 ◦ g1 ) = G(g2 ) ◦ G(g1 ) by showing that the morphism

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

165

G(g2 ) ◦ G(g1 ) satisfies the defining property of G(g2 ◦ g1 ). Indeed, using the fact that F is a functor and the defining properties of G(g1 ) and G(g2 ), we get β B3 F (G(g2 )G(g1 )) = β B3 FG(g2 )FG(g1 ) = g2 β B2 FG(g1 ) = g2 g1 β B1 . It is immediate from the definition of G that β : FG ⇒ 1D is a natural transformation, and thus a natural isomorphism. Finally, we construct a natural isomorphism α : 1C ⇒ GF. For any object A in C, let α A be the unique morphism A → GF ( A) such that F (α A ) = β F−1A. We leave it as a highly instructive exercise to the reader to check that α is a natural isomorphism (see Exercise 5.2.6 for hints).  Example 5.14 All the functors in Example 5.9 are part of a dual equivalence. Indeed, consider again the functor D : Pos f → (DL f ) op from Example 5.9(a). We check explicitly that D is full, faithful, and essentially surjective. It was shown in Proposition 1.21 of Chapter 1 that every finite distributive lattice is isomorphic to some lattice of the form D(P), where we can take for P the finite poset of join-prime elements of the lattice. Thus, D is essentially surjective. It was shown in Proposition 1.24 that, for arbitrary finite posets P, Q, the assignment f → D ( f ) := f −1 is a bijection between the set of order-preserving functions from P to Q and the set of lattice homomorphisms from D(Q) to D(P). This means exactly that D is full and faithful. Therefore, by Theorem 5.13, D is part of a dual equivalence. The functor J : (DL f ) op → Pos f gives the (up to natural isomorphism) unique functor in the other direction. To see this, one may either trace the proof of Theorem 5.13 in this specific case and check that J works as a choice for G in that proof, or directly exhibit the required natural isomorphisms. A natural isomorphism α : 1(DL f )op ⇒  : L → D(J (L)), which are isomorphisms D J is given by the family of maps (−) according to Proposition 1.21. A natural isomorphism β : 1Pos f ⇒ J D is given by the family of isomorphisms β P : P → J (D(P)) that send p ∈ P to ↓p, given by Proposition 1.22. This concludes our detailed proof of Theorem 1.25. You are asked to supply a similar proof of the Priestley duality theorem, Theorem 3.31, via Theorem 5.13 in Exercise 5.2.7, also see Theorem 5.38 for a direct proof that does not use Theorem 5.13. Note that, once the functors have been defined, these proofs always follow the same pattern, and have very little to do with the precise categories at hand. The real work in establishing a (dual) equivalence between two specific categories is in defining a full, faithful, essentially surjective functor from one category to another. The rest is, as some category theorists like to call it, “abstract nonsense.” In the next subsection, we will encounter some pairs of functors that do not form equivalences, but that do enjoy a looser bond.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

166

5 Categorical Duality

Adjunctions An important weakening of the notion of equivalence is that of an adjunction between categories. We already encountered adjunctions in the context of preordered sets in the very first section of this book (Definition 1.14). We used adjunctions between posets several times in Section 4.2, to formulate the correspondence between quotient lattices and subspaces. Adjunctions between categories generalize adjunctions between preorders, if we view a preorder as a category with at most one morphism between two objects (see Example 5.3). We have already seen examples of such categorical adjunctions “in action:” both the Boolean envelope (Section 3.3) and the free distributive lattice (Section 4.1) are examples of adjoint constructions; see Example 5.17 for details. Definition 5.15 Let C and D be categories and let F : C  D : G be a pair of functors between them. We say that (F, G) is an adjunction if there exist natural transformations η : 1C → GF and  : FG → 1D satisfying the following two properties: (a)

for any object A of C,  F A ◦ F (η A ) = 1F A; and

(b)

for any object B of D, G( B ) ◦ η GB = 1GB .

In this situation, F is called left or lower adjoint to G and G is called right or upper adjoint to F; a common notation for this is F & G. When the natural transformations η and  are explicitly specified, the natural transformation η is called the unit of the adjunction and  is called the co-unit of the adjunction. Properties (a) and (b) in Definition 5.15 are called triangle identities, as they can be expressed by saying that the following two triangles of natural transformations commute: F

Fη

FGF

G

F

1F F

ηG

1G

GFG G G.

While Definition 5.15 does not make the unit–co-unit pair (η,  ) part of the structure of an adjunction, and there may exist distinct choices of unit and co-unit for the same pair of functors, in the examples we consider in this book, there will always be a natural choice for η and , that we will specify when needed. An adjoint pair of functors is an equivalence exactly if the unit and co-unit can be chosen to be

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

167

natural isomorphisms, in the sense that, for any equivalence, there exist natural isomorphisms that satisfy the triangle identities (see Exercise 5.2.8). 2 The reader may recall that, for preorders P and Q, two order-preserving functions f : P  Q : g form an adjunction if, for any p ∈ P and q ∈ Q, f (p) Q q if, and only if, p  P g(q). An equivalent definition, given in Exercise 1.1.8, says that the function 1 P is pointwise below the function g f , and that f g is pointwise below 1Q , which shows why the notion of adjunction between categories generalizes that between preorders. In the setting of categories, we have the following statement, which is closer to the original definition of adjunction between preorders. Proposition 5.16 If F : C  D : G is an adjunction, then there exists, for any objects A in C and B in D, a bijection between the sets HomD (F A, B) and HomC ( A, GB). In other words, the cardinality of the set of morphisms from F A to B in D is equal to that of the set of morphisms from A to GB in C. Since, in a preorder, the number of morphisms between two objects is at most one, the reader should now clearly see the correspondence with the preorder definition of adjunction. We note here that Proposition 5.16 can also lead to an alternative definition of adjunction between categories, by adding the information that the family of bijections HomD (F A, B) → HomC ( A, GB) in fact forms a natural isomorphism between HomD (F−, −) and HomC (−, G−), when these assignments are viewed as functors from Cop × D to Set. Proving that this definition is equivalent to our Definition 5.15 is beyond the scope of this book, and we will work with Definition 5.15 as the official definition of adjunction. For a treatment of adjunctions that takes the Hom-set definition as primitive, see, for example, Mac Lane (1971, Chapter 4). An important property of adjunctions is that they compose: When F : C  D : G and F  : D  E : G  are functors such that F is left adjoint to G and F  is left adjoint to G , then F  F is left adjoint to GG  (see Exercise 5.2.13). Example 5.17 (a) Let U : DL → Set denote the forgetful functor of Example 5.9(c). The left adjoint to this forgetful functor is the free distributive lattice functor, FDL : Set → DL, whose object part was constructed in Section 4.1. The universal property allows one to extend this object assignment to a functor which is left adjoint to U. You are asked to supply the details of this example in Exercise 5.2.9. 2

There is a slight subtlety here: For an equivalence, not every pair of natural isomorphisms 1C → GF and FG → 1D satisfies the triangle identities, but there is always some choice that does. An equivalence together with explicit natural isomorphisms that satisfy the triangle identities is sometimes called an adjoint equivalence in the literature, to distinguish it from the non-adjoint situation, but for all of the (dual) equivalences in this book there is a straightforward choice of natural isomorphisms that do satisfy the triangle identities, so we need not make this distinction.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

168

(b)

(c)

(d)

(e)

5 Categorical Duality

Similarly to the previous item, the Boolean envelope of a distributive lattice (see Section 3.3) is the image of the left adjoint (−) − : DL → BA to the forgetful functor U : BA → DL. This item, together with the previous one and the fact that adjunctions compose, gives a more abstract proof of Lemma 4.11: The composition of (−) − and FDL gives a left adjoint FBA : Set → BA to the forgetful functor U : BA → Set. The free join-semilattice over a poset P was given in Exercise 1.3.8 as the lattice of finitely generated down-sets of P. This can be used to show that the forgetful functor U : SL → Pos, where SL is the category of semilattices, has a left adjoint, Dfin : Pos → SL. Let DL+ denote the (non-full) subcategory of distributive lattices that are isomorphic to a lattice of the form D(P), for P a poset, with complete homomorphisms between them. (The class of distributive lattices that are isomorphic to the down-set lattice of some poset may be characterized as “completely distributive complete lattices that have enough completely join-irreducible elements.” Completely distributive complete lattices will be studied extensively in Section 7.1.) There is a (non-full) inclusion functor I : DL+ → DL. The functor I has a left adjoint, () δ , which may be constructed as follows. For any distributive lattice L, let L δ be the lattice D(X L ), where X L is the poset underlying the Priestley dual space of L. For any distributive lattice homomorphism h : L → M, let h δ be the complete homomorphism ( f h ) −1 : L δ → M δ , where f h : X M → X L is the (order-preserving) function dual to h. The functor () δ is (naturally isomorphic to) the so-called canonical extension functor. The unit of the adjunction I & () δ is the function η L : L → D(X L ) that sends a ∈ L to a. ˆ The co-unit of the adjunction is defined, for C an object of DL+ , as the map  C : D(XC ) → C which sends a down-set D ⊆ XC to   Fx , where the join and meet are taken in the complete lattice C. x ∈D Another basic example of a free construction, which will become relevant in Chapter 8 of this book, is the free monoid on an input alphabet. A monoid is a tuple (M, ·, 1) where · is an associative binary operation on M, that is, u · (v · w) = (u · v) · w for all u, v, w ∈ M, and 1 is a neutral element, that is, 1 · u = u = u · 1 for all u ∈ M. In this context, a monoid homomorphism is a function f : M → N that preserves the multiplication and the identity element. We denote the category whose objects are monoids and whose morphisms are monoid homomorphisms by Mon. Note that this category Mon is different from the category associated to a single monoid M that we saw in Example 5.3(b). We have an obvious forgetful functor U : Mon → Set. This functor has a left adjoint, which we denote (−) ∗ : Set → Mon, and which sends a set X to the collection of finite sequences over X, equipped with multiplication given by concatenation of sequences, for which the empty sequence is an

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

169

identity element. Exercise 5.2.10 asks you to show that this indeed yields a left adjoint to the forgetful functor. The similarity between items (e) and (a) in Example 5.17 is no coincidence: General results of universal algebra, going back to Birkhoff, show that any socalled variety of algebras admits a free construction (see Exercise 5.2.11). In fact, any adjunction between categories may be understood as “a universal construction,” in the following sense. Theorem 5.18 Let G : D → C be a functor. The following are equivalent: (i) (ii)

The functor G has a left adjoint. For every object A of C, there exist an object F ( A) of D and a morphism η A : A → G(F ( A)) of C such that, for any object B of D and any Cmorphism f : A → G(B), there is a unique D-morphism f¯ : F ( A) → B such that (G f¯) ◦ η A = f .

This theorem generalizes the characterization of functors that are part of an equivalence, Theorem 5.13, in the following sense: If a functor G is essentially surjective, then one may always choose an object F ( A) and an isomorphism η A : A → G(F ( A)); if G is then moreover full and faithful, then any such choice will satisfy the universal property stated in (ii) of Theorem 5.18, and will thus have a left adjoint, with which G will then form an equivalence. The proof of Theorem 5.18 is beyond the scope of this book; see for example Borceux (1994, Section 3.1). We end our discussion of adjunctions by connecting it to equivalences in a different useful way: One may always obtain a (dual) equivalence from a (contravariant) adjunction. Indeed, Theorem 5.13 can be applied to prove that any (contravariant) adjunction restricts to a maximal (dual) equivalence. This associated duality is obtained by restricting the functors to the full subcategories given by the objects for which the natural map between the object and its double dual is an isomorphism. We record this fact here and leave the precise proof as Exercise 5.2.15. Corollary 5.19 Let F : C  D : G be an adjunction. Denote by Fix(C) the full subcategory of C consisting of those objects A for which the unit η A : A → GF A is an isomorphism. Similarly, denote by Fix(D) the full subcategory of D consisting of those objects B for which the co-unit  B : FGB → B is an isomorphism. Then, F and G restrict to an equivalence between Fix(C) and Fix(D).

Isomorphisms Finally, the notion of isomorphism between categories is easy to state, but often too restrictive.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

170

5 Categorical Duality

Definition 5.20 A functor F : C → D is called an isomorphism between categories C and D if there exists a functor G : D → C such that G ◦ F is equal to the identity functor on C and F ◦ G is equal to the identity functor on D. A dual isomorphism between C and D is an isomorphism between C and Dop . Two categories C and D are called isomorphic if there exists an isomorphism between them and dually isomorphic if there exists a dual isomorphism between them. Note that an isomorphism between locally small categories is, by definition, a functor that is an isomorphism, in the sense of Definition 5.4, when viewed as a morphism in the (large) category of locally small categories. Example 5.21 As indicated in Exercise 2.2.6, the category Top f of finite T0 topological spaces is isomorphic to the category Pos f of finite partially ordered sets. Indeed, for a finite T0 space (X, τ), let F (X, τ) := (X, ≤τ ) be the specialization order of (X, τ), and, conversely, for a finite order (X, ≤), let G(X, ≤) := (X, α(X )) be the set X equipped with the Alexandrov topology induced by ≤. These object assignments F and G extend to functors by sending any function to itself. Example 5.22 In Section 2.3 in Chapter 2, we studied a bijective correspondence between compact ordered spaces and stably compact spaces: To a compact ordered space (X, τ, ≤), assign the stably compact space (X, τ ↑ ), and conversely to a stably compact space (X, σ), assign the compact ordered space (X, σ p, ≤σ ). We saw in Proposition 2.20 that these assignments are mutually inverse. Compact ordered spaces form a category whose morphisms are the continuous order-preserving maps. If we equip the class of stably compact spaces with the proper maps defined in Exercise 2.3.15, then it will become isomorphic to the category of compact ordered spaces; you will be asked to supply the details in Exercise 5.2.16. In the next chapter, Section 6.1, we will study the restriction of this isomorphism of categories to the full subcategory of Priestley spaces; the image on the side of stably compact spaces is the class of spectral spaces.

Exercises for Section 5.2 Exercise 5.2.1 This exercise guides you through the details of identifying Pos f with a full subcategory of Priestley. (a) (b) (c)

Prove that the topology on a finite compact Hausdorff space is discrete. Prove that, if (X, ≤) is a finite poset, then ι(X, ≤) := (X, τ, ≤), where τ is the discrete topology on X, is a Priestley space. Prove that a function f : X → Y between finite Priestley spaces is a morphism in Priestley if, and only if, f is order preserving.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.2 Constructions on Categories

(d)

171

Conclude that the object assignment ι in item (b) extends to a full and faithful functor ι : Pos f → Priestley.

Exercise 5.2.2 A category is called self-dual if C is equivalent to Cop . (a)

(b)

(c)

(d) (e) (f)

Verify that the category Rel is self-dual, and in fact, isomorphic to the category Relop . We now outline a proof that Set is not self-dual. Exhibit an object 1 in Set with the property that for any object X there is exactly one morphism X → 1. (Such an object is called a terminal object in a category, and is unique up to isomorphism, as we will see in Section 5.3.) Show that Set has the following property: For any two distinct morphisms f , g : X → Y , there exists a morphism x : 1 → X such that f x  f y. A category with this property is said to be generated by its terminal object or well pointed. What is the terminal object in Setop ? Prove that Setop is not generated by its terminal object. Conclude that Set is not self-dual.

Exercise 5.2.3 This exercise outlines a possible construction of a skeleton of a category. The special case of preorders was already given in Exercise 1.1.5. Let C be a category. Choose a collection of objects S in C such that every object in C is isomorphic to exactly one object of S; note that this step uses the axiom of choice for the collection ob C, which is in general not a set, but we ignore set-theoretic issues here. Prove that the inclusion functor of the full subcategory S on the collection of objects S is an equivalence. Exercise 5.2.4 Let F, G, H : C → D be functors and let ϕ : F ⇒ G and ψ : G ⇒ H be natural transformations. (a) (b) (c) (d) (e) (f)

Give a definition of the composition ψ ◦ ϕ : F ⇒ H, and show that it is again a natural transformation. Show that there is an identity natural transformation 1F : F ⇒ F. Given a functor K : D → E, show that K ϕ : K F ⇒ KG, defined at an object C of C by (K ϕ) A := K (ϕ A ), is a natural transformation. Given a functor L : B → C, show that ϕ L : F L ⇒ GL, defined at an object B of B by (ϕ L )B := ϕ LB , is a natural transformation. Prove that, if ϑ : G → F is a natural transformation such that ϑ ◦ ϕ = 1F and ϕ ◦ ϑ = 1G , then ϕ and ϑ are natural isomorphisms. Prove that, if ϕ is a natural isomorphism, then defining ϑ A := (ϕ A ) −1 for every A ∈ ob C yields a natural transformation ϑ that is a two-sided inverse for ϕ.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

172

5 Categorical Duality

Exercise 5.2.5 Let F : C  D : G be an equivalence of categories. Prove that the functors F and G are full, faithful, and essentially surjective. Exercise 5.2.6 This exercise asks you to complete the proof of the sufficiency direction of Theorem 5.13. Consider the family of morphisms (α A ) A∈ob C defined at the end of the proof of that theorem in the text. (a)

Prove that α : 1C ⇒ GF is a natural transformation. Hint. For a morphism f : A → A in C, show first the equality β F A ◦ F (α A ◦ f ) = β F A ◦ F (GF ( f ) ◦ α A ),

(b)

and then explain why this is enough to conclude that α A f = GF ( f )α A. Prove that α A is an isomorphism for every object A. Hint. Consider a morphism γ A : GF A → A such that Fγ A = β F A.

Exercise 5.2.7 Use Theorem 5.13 to show that the results from Chapter 3 imply that the functors from Example 5.9(c) form a dual equivalence. In other words, prove Theorem 3.31 (Priestley duality). Note. We will give more details and in particular a somewhat more explicit proof of Priestley duality in Section 5.4. Exercise 5.2.8 Show that every equivalence can be improved to an adjoint equivalence. That is, let F : C  D : G be an equivalence of categories. Prove that there exist natural isomorphisms η : 1C → GF and  : FG → 1D satisfying the triangle identities. Hint. Consider the proof of Theorem 5.13. Exercise 5.2.9 This exercise asks you to supply the details omitted in Example 5.17(a) about the free distributive lattice on a set. For any set X, denote by η X : X → U FDL (X ) the embedding of X into the free distributive lattice over X. (a) (b) (c)

Prove that FDL extends to a functor Set → DL. Explicitly define a co-unit  : FDLU → 1DL . Prove that η and  are natural transformations satisfying the triangle identities.

Exercise 5.2.10 Repeat Exercise 5.2.9 for the free monoid functor from Example 5.17(e) (see Exercise 5.2.11 for a general result of which this exercise and the previous one are specific instances). Exercise 5.2.11 This exercise generalizes Exercises 5.2.9 and 5.2.10. Let V be a variety of finitary algebraic structures with at least one constant symbol in the signature. Denote by V the category whose objects are the algebras in V, and whose morphisms are the homomorphisms, and denote by U : V → Set the forgetful functor. Prove that U has a left adjoint FV : Set → V. Hint. This exercise essentially asks to prove a well-known theorem of Birkhoff, namely, that any finitary variety

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

173

has free algebras. The proof mimics exactly the algebraic construction of the free distributive lattice given in Section 4.1. Also see Burris and Sankappanavar (2000, Chapter II). The reason for assuming that the signature has a constant symbol is that, if the signature has no constant symbol, then FV (∅) has to be empty, and some authors do not allow the empty set as an algebra. Exercise 5.2.12 Prove Proposition 5.16. Exercise 5.2.13 Let F : C  D : G and F  : D  E : G  be adjunctions. Prove that F  F : C  D : GG  is an adjunction. Exercise 5.2.14 Prove that “equivalence” is an equivalence relation on categories. That is, show that any category C is equivalent to itself, that C equivalent to D implies D equivalent to C, and that if C is equivalent to D and D is equivalent to E, then C is equivalent to E. Exercise 5.2.15 Show that Corollary 5.19 indeed follows from Theorem 5.13. Exercise 5.2.16 Using the results of Exercise 2.3.15, prove that the bijective correspondence of Theorem 2.21 extends to an isomorphism of categories, as outlined in Example 5.22 (also see Theorem 6.4 in Chapter 6).

5.3 Constructions inside Categories We now come to the internal perspective on a category. We will consider many types of internal notions and constructions in a category: monomorphisms and epimorphisms; products and coproducts; equalizers and coequalizers; and, most generally, limits and colimits. The reader will notice that these notions come in pairs; why? As we discussed early on in Section 5.2, any notion in a category C can also be considered in the so-called opposite category Cop , in which the direction of the morphisms is reversed. This means that, for any concept of a categorical “gadget,” one can also define a “co-gadget” in C to be “a thing in C that is a gadget when viewed in Cop .” The name for the dual notion is often formed by prefixing “co-” to the original name: product and coproduct; equalizer and coequalizer; limit and colimit. 3 In this section, we often explicitly introduce both a notion and its dual. While not the most economical, this will allow the reader to see both the formal similarity and the fact that these notions can have very different concretizations in specific examples of categories. 3

A notable exception to this terminological rule is “covariant” vs “contravariant” functor, defined in the previous section, and also “monomorphism” vs. “epimorphism” introduced here.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

174

5 Categorical Duality

Recall our convention that we often omit the symbol ◦ for composition within a category, and we write strings like “h g f ” as shorthand for “h ◦ g ◦ f .” Monomorphisms, Epimorphisms, and Coequalizers We already saw one type of special morphism in a category: an isomorphism, defined in Definition 5.4. We now introduce two related, but weaker, notions. Definition 5.23 Let f : A → B be a morphism in a category C. We say that f is a • monomorphism if, for any g, h : A → A in C such that f g = f h, we have g = h, and an • epimorphism if, for any g, h : B → B  in C such that g f = h f , we have g = h. We often use the abbreviations “mono” and “epi,” omitting the suffix “-morphism.” It is not hard to prove that in any category, an iso is both mono and epi. The converse, however, fails in many categories (see Exercises 5.3.1 and 5.3.4 for examples). In the category of sets, monomorphisms are just injective functions and epimorphisms are surjective functions. However, this correspondence breaks down in most other concrete categories! In particular, categories of algebras often have non-surjective epimorphisms; for an example, see Exercise 5.3.1 for the case of distributive lattices. An abstract notion that corresponds more closely to “surjective,” at least in categories of algebras, is that of a regular epimorphism, which we introduce now. Definition 5.24 Let p1, p2 : C → A and q : A → Q be morphisms in a category C. We say that q is a coequalizer of (p1, p2 ) if (i) qp1 = qp2 and (ii) for any morphism q  : A → Q  in C such that q  p1 = q  p2 , there exists a unique morphism α : Q → Q  in C such that αq = q . A regular epimorphism q : A → Q in C is a morphism which is the coequalizer of some pair of morphisms. The terminology suggests that any morphism which arises as a coequalizer is in fact an epimorphism; you are asked to prove that this is indeed the case in Exercise 5.3.5. In the following example, we explicitly compute coequalizers in the category Set. Example 5.25 Let p1, p2 : C → A be functions between sets. Define ≡ to be the smallest equivalence relation on the set A that contains the pair (p1 (c), p2 (c)), for every c ∈ C. Let Q := A/≡ be the quotient and q : A → Q the quotient map. We show that q is a coequalizer of (p1, p2 ). Clearly, qp1 = qp2 , since ≡ contains all pairs (p1 (c), p2 (c)). Suppose that q  : A → Q  is any function such that q  p1 = q  p2 . Note that the equivalence relation ker(q  ) = {(a, b) ∈ A2 | q  (a) = q  (b)} contains every pair (p1 (c), p2 (c)), and therefore must contain the equivalence relation ≡, since

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

175

this is the smallest such. Therefore, there is a well-defined and unique factorization α : Q → Q  such that αq = q . Exercise 5.3.7 outlines a proof that in the category of lattices, the regular epimorphisms coincide with the surjective homomorphisms. Dually, we have notions of equalizer and regular monomorphism, that we will discuss in more detail below. In the category of lattices, a lattice homomorphism is a monomorphism if, and only if, it is injective. Not every monomorphism is regular in the category of distributive lattices (see Exercise 5.3.1). The definition of coequalizer is an example of a universal property: It states the existence of a “minimal solution” for a configuration of morphisms in a category; in this case, a coequalizer q of (p1, p2 ) “solves for x” the equation xp1 = xp2 , and does so in a “minimal” way, in the sense that any other solution q  factors through q. A coequalizer is a special case of a colimit in a category, as we will explain below. Any concept defined by a universal property is unique up to unique isomorphism. Concretely, for coequalizers this means that if q1 and q2 are both coequalizers of (p1, p2 ), then the unique maps between the codomains of q1 and q2 which are guaranteed to exist by the definition must be inverse to each other (see Exercise 5.3.6). We will see several more examples of concepts defined by universal properties in this section. In fact, universal properties, limits, and adjunctions are all closely related concepts; the interested reader may find, after digesting the definition of limit, an indication of the connection in Exercise 5.3.14.

Products and Coproducts As we saw in Example 5.25, coequalizers are an appropriate categorical generalization of “quotient map.” The Cartesian product and disjoint union of sets admit a similar generalization, namely as product and coproduct, respectively. Definition 5.26 Let I be a set and (Di )i ∈I an I-indexed family of objects in a category C. A product of the family (Di )i ∈I is an object L together with an Iindexed family of morphisms (πi : L → Di )i ∈I such that, for any object A and any I-indexed family of morphisms (ai : A → Di )i ∈I , there exists a unique morphism α : A → L such that πi ◦ α = ai for every i ∈ I.  If a family has a product, we often denote it by i ∈I Di , and, if I = {1, 2}, by D1 × D2 . In the case I = ∅, a product of the unique ∅-indexed family is called a terminal object, and is often denoted by 1. Note that an object T in a category C is terminal if, and only if, for every object A in C, there is a unique morphism from A to T. When a1 : A → D1 and a2 : A → D2

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

176

5 Categorical Duality A a1, a2 

a1

a2

D1 × D2 π1

π2

D1

D2

Figure 5.1 The universal property of binary product in a category.

are morphisms, the unique morphism A → D1 × D2 is often denoted a1, a2 . The universal property can then be expressed in a diagram, as in Figure 5.1. Also, when f 1 : A1 → D1 and f 2 : A2 → D2 are morphisms, then there is a unique morphism f : A1 × A2 → D1 × D2 defined by  f 1 ◦ π1, f 2 ◦ π2 . This morphism f is denoted f 1 × f 2. It is easy to verify that the Cartesian product of sets gives a product in Set. Any one-element set is a terminal object in Set. Products exist in any category of finitary algebraic structures and are given by the Cartesian product in such a category; in particular, the categorical product in categories of lattices is given as in Set (see Exercise 5.3.9). The product in Priestley spaces is also given as in the category of ordered topological spaces, because the product of a collection of Priestley spaces is a Priestley space (see Exercise 5.3.10). If a category has finite products, then this allows us to “internalize” algebraic structure in the category; this perspective was popularized by Lawvere, giving rise to a notion of Lawvere theory. While we do not need to develop this theory here, we give an example that will be relevant in Chapter 8 and that gives the flavor of the idea. Example 5.27 A monoid internal to a category C is an object M of C, together with two morphisms I : 1 → M and μ : M × M → M such that the following three diagrams commute: 1×M

I ×id M π2

M×M

M ×1

μ

M

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

id M ×I π1

M×M μ

M

5.3 Constructions inside Categories μ×id M

(M × M) × M

M×M μ

α

M × (M × M)

177

id M ×μ

M×M

μ

M,

where, in the last diagram, the morphism α is the associator defined by  f 1, f 2 , where f 1 : (M × M) × M → M is the composition of the two projections on the first coordinate and f 2 : (M × M) × M → M × M is π2 × id M . These diagrams correspond to the monoid axioms 1 · a = a, a · 1 = a, and (a · b) · c = a · (b · c), respectively. For example, a topological monoid, which we will encounter again in Definition 8.30, is a monoid internal to Top and a Priestley monoid is a monoid internal to Priestley. For these concrete categories, it is easy to see that, for instance, a topological monoid (M, I, μ) is the same thing as a monoid on the underlying sets for which μ is continuous, and, in the case of Priestley, order preserving (the function I always is because the object 1 is a one-point space). An example of a different flavor is a monad, which can be seen as a monoid internal to the category of endofunctors; however, properly defining monads in this way requires replacing in the above definition the product × by the composition of endofunctors, known as monoidal structure on a category. We now give the dual definition of coproduct. Definition 5.28 Let I be a set and (Di )i ∈I an I-indexed family of objects in a category C. A coproduct of the family (Di )i ∈I is an object C together with an I-indexed family of morphisms (ιi : Di → C)i ∈I such that, for any object B and any I-indexed family of morphisms (bi : Di → B)i ∈I , there exists a unique morphism β : C → B such that β ◦ ιi = bi for every i ∈ I.  If a family has a coproduct, we often denote it by i ∈I Di , and, if I = {1, 2}, by D1 + D2 . In the case I = ∅, a coproduct of the unique ∅-indexed family is called an initial object, and is often denoted by 0. Remark 5.29 We occasionally used the notation A + B earlier in this book for the symmetric difference of two sets A and B, usually as subsets of a common set X as, for example, in Example 4.21. Note that this symmetric difference A + B is not in general a coproduct of the objects A and B in the category Set, but it is when the sets A and B are disjoint. An initial object in a category is an object 0 such that there exists a unique morphism from 0 to any object of the category. Coproducts in Set are given by disjoint unions and the initial object is the empty set. However, coproducts are often more complicated to compute in other categories of algebras and spaces. For any

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

178

5 Categorical Duality



set V , v ∈V 3 is the free distributive lattice over the set of generators V , since 3 is the free distributive lattice on a single generator (see Exercise 5.3.13). Finite coproducts in Priestley spaces are disjoint unions, but infinite coproducts involve a compactification step (see Exercise 5.3.11). Example 5.30 Let P be a preorder, viewed as a category, and let (pi )i ∈I be a family of objects in P. A product of (pi )i ∈I is exactly an infimum and a coproduct is exactly a supremum. A terminal object in the category P is a greatest element and an initial object is a least element. Note that the convention of drawing a product above the objects (as in Figure 5.1) clashes with the convention of Hasse diagrams (as, for example, in Figure 1.1), in which an object q of P is depicted above an object p if p ≤ q. As we have seen in Chapter 1, the value of infima and suprema depend on the ambient poset they are taken in. Thus, the example above illustrates clearly that products and coproducts depend not only on the objects in question, but also on the ambient category the product or coproduct is taken in. Limits and Colimits In our definition of limit and colimit below (Definition 5.32), the following formal notion of diagram plays a central role. We have been drawing diagrams informally throughout this book, but a diagram can itself be seen as a categorical object. We first give the formal definition. Definition 5.31 Let C be a category and I a small category. An I-shaped diagram in C is a functor D : I → C. As a motivating example for this formal definition, which can look surprising at first, let us see how we would describe the following diagram (Figure 5.2) according to Definition 5.31. D1 Dx 13 D3

Dx 12 Dx 14

Dx 34

D2 Dx 24 D4

Figure 5.2 An example of a commutative diagram, formally.

Let I be the category defined by ob I := {1, 2, 3, 4}, with the following five non-identity morphisms: x 12 : 1 → 2, x 13 : 1 → 3, x 24 : 2 → 4, x 34 : 3 → 4,

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

179

and x 14 , which is by definition equal to both x 24 ◦ x 12 and x 34 ◦ x 13 . This indeed defines a category I; associativity is trivial because there is no way to compose three non-identity morphisms in this category. Now, note that a functor D : I → C is essentially the same thing as a commutative square in C as depicted in Figure 5.2, where we note that the value of Dx 14 is determined by functoriality of D and the equalities x 24 ◦ x 12 = x 14 = x 34 ◦ x 13 holding in I. Indeed, we have Dx 14 = D(x 24 ◦ x 12 ) = Dx 24 ◦ Dx 12 , and similarly Dx 14 = Dx 34 ◦ Dx 13 , showing that the diagram commutes. We are now ready to define limits and colimits. Two instances of this very general concept that can be helpful to keep in mind while reading this definition are the following: • Let I be a set and let I be the small category with ob I = I and with only identity morphisms, no morphisms between distinct objects – this is called the discrete category on I. An I-shaped diagram in C is just given by an I-indexed family of objects in C. A product as defined in Definition 5.26 is a limit of an I-shaped diagram and a coproduct is a colimit of an I-shaped diagram. • Let J be the small category with two objects, 1 and 2, and two morphisms from 1 to 2. A J-shaped diagram in a category C is a pair of objects in C with a parallel pair of morphisms between them. A colimit of a J-shaped diagram is a coequalizer as defined in Definition 5.24 and a limit of such a diagram is called an equalizer. Spelling out the definition of the latter, an equalizer of a pair of parallel arrows f 1, f 2 : A ⇒ B in a category C is an object E and a morphism e : E → A such that f 1 e = f 2 e and, for any morphism a : C → A such that f 1 a = f 2 a, there exists a unique α : C → E such that a = eα. We now give the general definition of limit and colimit of a diagram in a category. Definition 5.32 Let C be a category, let I be a small category, and let D : I → C be an I-shaped diagram in C. A cone above the diagram D is an object A together with an ob I-indexed family of morphisms in C, (ai : A → Di)i ∈ob I , such that, for any morphism x : i → j in I, we have (Dx) ◦ ai = a j . A limit of the diagram D is a universal cone above D. That is, a limit of D is an object L with an ob I-indexed family of morphisms, (πi : L → Di)i ∈ob I , such that, for any cone (ai : A → Di)i ∈ob I above D, there exists a unique morphism α : A → L in C such that πi α = ai for every i ∈ ob I. Similarly, a cone below, or also co-cone under, D is an object B together with morphisms bi : Di → B for every i ∈ ob I such that, for any x : i → j in I, b j ◦ Dx = bi . A colimit is a universal cone below D, that is, an object C together with morphisms ιi : Di → C for every i ∈ ob I such that, for any cone (bi : Di → B)i ∈ob I below D, there exists a unique β : C → B such that βιi = bi for every i ∈ ob I.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

180

5 Categorical Duality

We note here that a cone above a diagram D can be viewed more conceptually as a natural transformation from a constant functor Δ to D and a cone below D as a natural transformation from D to a constant functor (see Exercise 5.3.8 for more details). Using this observation, one may see a limit of D as a terminal object in a category of cones over D. As we allowed the shape of the indexing category of the diagram to be any small category in Definition 5.31, the reader may reasonably worry that limits and colimits could have incredibly complicated shapes. While this is true, the relevant shapes that we need to consider in practice are rather limited, thanks to a few bits of general category theory that we will develop in Proposition 5.37. In particular, we will explain why it is enough for our purposes to understand the following three types of limits/colimits: • product and coproduct; • equalizer and coequalizer; and • projective limit and directed colimit. Projective Limits and Directed Colimits The definitions of a projective limit and directed colimit use the notion of directed poset: Recall that a poset P is directed (sometimes called up-directed) if it is nonempty and any pair of elements p, q has a common upper bound, and down-directed if Pop is up-directed (see p. 13). Definition 5.33 A diagram D : I → C is directed if I is the category associated to an up-directed partial order and projective if I is the category associated to a down-directed partial order. A projective limit is a limit of a projective diagram; a directed colimit is a colimit of a directed diagram. 4 Some sources are more liberal in their definitions of directed colimit and projective limit, by allowing the shape of the diagram I to be a directed preorder rather than partial order. However, this does not give a more general notion of limit, up to isomorphism (see Exercise 5.3.15). A seemingly even more general definition is that of a colimit with respect to a filtered category, that is, a non-empty category in which any pair of objects admits a morphism to a common object and any parallel pair of morphisms has a cone under it. Surprisingly, colimits with respect to filtered categories are no more general than colimits with respect to directed posets; the proof of this fact is beyond the scope of this book, see, for example, Adámek and Rosický (1994, Theorem 1.5). 4

In algebraic literature, what we call “directed colimit” is sometimes called “direct limit” or “inductive limit.” We avoid this terminology because we prefer to emphasize “co-” in the name of this concept, to signal the fact that this generalizes a “coproduct.” On the other hand, in category theory, what we call “projective limit” is sometimes called “down-directed limit” or “codirected limit.”

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

181

In a concrete category, projective limits and directed colimits of finite-like objects are often particularly interesting to look at as they tend to inherit some of the properties of finite-like objects. To give one pertinent class of examples, in the category of algebras of a finitary variety (see Exercise 5.2.11), a reasonable notion of “finite-like object” is that of a finitely presented algebra. Here, an algebra is called finitely presented if it is a quotient of a finitely generated free algebra by a finitely generated congruence. Then, every algebra is a directed colimit of finitely presented algebras. This is essentially the defining property of what is known as an ℵ0 accessible category in the literature, see for example Adámek and Rosický (1994). In the particular case of distributive lattices, the situation simplifies further since any finitely presented distributive lattice is in fact finite, as follows immediately from the fact that free distributive lattices are finite. Thus, every distributive lattice is a directed colimit of finite distributive lattices. We now sketch a direct proof of the latter fact, which should give an impression of the basics of the much more general theory. Example 5.34 Let L be a distributive lattice. Let Sub f (L) denote the poset of finite sublattices of L, which is up-directed, because if F and G are finite sublattices of L, then they are both contained in the sublattice generated by F ∪ G, which is still finite. Let D : Sub f (L) → DL be the diagram defined on objects by D(F) := F and sending any sublattice inclusion F ⊆ G to the homomorphism D(F ⊆ G) := i F,G : F → G. Note that a co-cone under this diagram D is given by the inclusion maps ιF : F → L. Indeed, this is a colimit: If ( j F : F → M)F ∈Sub f (L) is another co-cone under this diagram, then a unique α : L → M may be defined by sending a ∈ L to the element j L a (a), where L a denotes the (at most three-element) sublattice of L generated by a. To prove that α is indeed a homomorphism, one may use the fact that any pair of elements generates a finite sublattice of L, and that the j F form a co-cone. Generalizing this example slightly, let P be a directed subset of Sub f (L). One may then prove that the union of the finite sublattices in P is a sublattice L  of L and that L  is the directed colimit of the restricted diagram D| P . The directed colimit of P is (isomorphic to) L itself if, and only if, every a ∈ L lies in some sublattice in the collection P. This characterization of directed colimits will be used in Section 7.3 to give a concrete definition of bifinite distributive lattices. For an example of the importance of projective limits of finite objects, it is interesting to move to the topological side of Priestley duality. Indeed, Priestley spaces may be characterized as the profinite posets, in a sense that we make precise now in the following example. Example 5.35 In the category TopOrd of ordered topological spaces with continuous order-preserving functions, we have a full subcategory isomorphic to the

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

182

5 Categorical Duality

category of finite posets. Indeed, for every finite poset P, we have an ordered topological space, that we still denote by P, by equipping P with the discrete topology. With this topology, P is a finite Priestley space, and any order-preserving function between finite posets is continuous for the discrete topology, as we saw in Exercise 5.2.1. We denote this category of finite posets with the discrete topology by TopOrd f , d . We now use Example 5.34 and Priestley duality to show that every Priestley space is a projective limit of finite Priestley spaces. Let X be a Priestley space and let L be its dual distributive lattice. Since L is the colimit of the diagram D : Sub f (L) → DL given in Example 5.34, it follows by a purely formal argument, since DLop is equivalent to Priestley, that X is the limit of the diagram D  : (Sub f (L)) op → Priestley that sends a finite sublattice F of L to the Priestley dual space of F. We may give a more direct description of the diagram D  using the sublattice-quotient duality of Section 4.2: The domain of D  is (isomorphic to) the poset of compatible preorders  on X for which X/ is finite and it sends such a preorder  to the finite Priestley space X/. Since finite Priestley spaces always have discrete topology, an equivalent way to describe D  is as the diagram of all continuous surjective maps X  P where P is a finite poset and X is the limit of this diagram in the category TopOrd. Also, since any limit of finite posets is a Priestley space, we conclude that Priestley spaces exactly coincide with the projective limits of finite posets in TopOrd. As a special case, Boolean spaces exactly coincide with the projective limits of finite sets with the discrete topology in Top, and are therefore sometimes called profinite sets. The “profinite” point of view on Stone–Priestley duality will play an important role in Chapter 8, where it is linked to the theory of profinite monoids. The categorical definition of profinite monoids is analogous to the ones in this example: The category of finite monoids is isomorphic to a full subcategory of the category TopMon topological monoids (Example 5.27) by equipping each finite monoid with the discrete topology. A profinite monoid is then an object of TopMon that is a projective limit of finite monoids. Further, see Definition 8.30.

Remark 5.36 We remark that the arguments given in Example 5.35 can be turned around to give a categorical proof of the Priestley duality theorem itself from Birkhoff duality for finite distributive lattices, as we now outline, without giving all the details. Example 5.34 shows, in categorical terms, that the category DL may be obtained by taking directed colimits of objects from the category DL f . Therefore, its (formal) dual category DLop can be described by taking projective limits of the category (DL f ) op . The latter category is equivalent to the category of finite posets, by Birkhoff duality, and this is isomorphic to the full subcategory TopOrd f ,d of TopOrd on the finite posets with the discrete topology. It follows

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

183

that DLop is equivalent to the closure of TopOrd f ,d under projective limits. This gives a “formal” version of Priestley duality. Note that we have not used any choice principles to deduce this duality, only Birkhoff duality and abstract category theory. In this point of view on Priestley duality, the choice principle needs to be used only in the very last step, namely to prove that the closure under projective limits of TopOrd f , d is indeed the same as the category of Priestley spaces. This requires in particular showing that any product of finite spaces is compact, which is a “finite” Tychonoff theorem, and is in fact equivalent over choice-free set theory to the Stone prime ideal theorem. To understand the importance of products, equalizers, and projective limits for understanding general limits, we now first look at a special, finite case, where just binary products and equalizers will suffice. A generalization of this result, also involving projective limits, will be proved after that, in Proposition 5.37. Let C be a category and consider two morphisms in C, f 1 : A1 → B and f 2 : A2 → B, with a common codomain B, as in the following diagram. A1 f1 A2

f2

B.

Such a pair of morphisms is called a co-span in category theory, and a limit of such a diagram is called a pullback; the dual notions are called span and pushout, respectively. Now, assume that binary products and equalizers exist in C. We show that the limit (pullback) of any such co-span can be constructed in C, as follows. First, let P = A1 × A2 be the product of A1 and A2 in C, and consider two morphisms g1, g2 : P ⇒ B defined by gi := f i πi for i = 1, 2, where πi : P → Ai is the projection morphism. Now, let e : E → P be the equalizer of g1 and g2 . We may now prove that the pair of maps p1, p2 : E → Ai , defined by pi := πi e for i = 1, 2, is a limit cone for the co-span. Indeed, using the above definitions, we have that f 1 p1 = f 1 π1 e = g1 e = g2 e = f 2 π2 e = f 2 p2, so that the pi form a cone over the diagram. Furthermore, if a1 : A → A1 and a2 : A → A2 are another cone over the same diagram, then, writing a for the morphism a1, a2  : A → A1 × A2 , we get g1 a = f 1 π1 a = f 1 a1 = f 2 a2 = f 2 π2 a = g2 a,

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

184

5 Categorical Duality

using the definitions and the fact that the ai form a cone. Thus, there exists a unique α : A → E such that eα = a. It follows that ai = πi a = πi eα = pi α for i = 1, 2. Also, such α is unique: If α  is such that a1 = p1 α  and a2 = p2 α , then πi eα  = pi α  = ai for i = 1, 2, so that eα  = eα by the universal property of the product P. Therefore, α  = α since equalizers are monomorphisms (see Exercise 5.3.5). A generalization of the above construction allows one to show that, if all equalizers and all set-sized products exist in a category, then the limit of any small diagram exists. We prove this fact now and further connect it to the construction of projective limits in a category. Proposition 5.37 Let C be a category. The following are equivalent: (i) (ii) (iii)

Any small diagram has a limit in C. Any finite diagram has a limit in C, and any projective diagram has a limit in C. All equalizers and small products exist in C.

Proof It is trivial that the first item implies the second. For the implication (ii) ⇒ (iii), since an equalizer is the limit of a finite diagram, it suffices to prove that any set-sized product exists in C. Let ( Ai )i ∈I be a family of objects of C indexed by a set I. We will show that a product of the Ai can be constructed as a limit of a diagram B of shape P f (I) op , the poset of finite subsets of I, ordered by reverse inclusion – this is clearly a down-directed poset, since it is a meet-semilattice, so the limit of B will exist by assumption. For each F ∈ P f (I), let B(F) be a product of the finite set of objects { Ai | i ∈ F}, which exists in C by assumption, and, for i ∈ F, denote by πiF : B(F) → Ai the projection of this cone on Ai . In the special case where F is a singleton, we choose B({i}) = Ai , which is the product of the single object Ai . If F ⊆ G, then we have a cone with summit B(G) over { Ai | i ∈ F}, since for each i ∈ F, we have i ∈ G, so we have πiG : B(G) → Ai . Let β F,G : B(G) → B(F) be the unique factorization of this cone, so that for every i ∈ F, πiF β F,G = πiG , and define B(G ⊇ F) := β F,G . Now, let L be a limit of the diagram B with λ F : L → B(F) the legs of the limit cone for each finite F ⊆ I. In particular, for every i ∈ I, write πi for the leg λ {i } : L → B({i}) = Ai . We claim that πi : L → Ai gives a product of the Ai . Indeed, if f i : A → Ai is any family of morphisms, then we may construct, for each finite subset F ⊆ I, a unique morphism f F : A → B(F) such that πiF f F = f i for each i ∈ F, using the universal property of the product B(F). The morphisms f F give a cone over the diagram B, because for any morphism G ⊇ F in P f (I) op , we have πiF β F,G f G = πiG f G = f i = πiF f F for every i ∈ F, so that βF,G f G = f F , using the uniqueness part of the universal property of the product B(F). The unique factorization α : A → L satisfies πi α = f i , as required.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

185

Finally, we sketch a proof that (iii) implies (i); we emphasize that this is an infinite generalization of the construction of pullbacks from binary products and equalizers, described just above this proposition. Let D : I → C be a diagram indexed by a small category I. Observe that giving a cone ai : A → Di over the diagram is the  same thing as giving a single morphism a : A → i ∈ob(I ) Di with the property that, for all morphisms f : i → j in I, A( f ) ◦ ai = a j . The latter property, in turn,  can be more synthetically expressed as follows. Write P for the product i ∈ob I Di, with projection maps πi : P → Di for every object i of I, and write M for the  product f ∈mor I D(cod( f )), with projection maps ρ f : M → D(cod( f )) for every morphism f of I. Now, the crucial observation is that a : A → P is a cone exactly if d ◦ a = e ◦ a, where d, e : P ⇒ Q are two morphisms of C defined, respectively, by requiring that for every morphism f of I: ρ f ◦ d := πcod( f ),

ρ f ◦ e := D( f ) ◦ πdom( f ) .

From this observation, it follows that the equalizer of the parallel pair of maps d, e is a limit cone of the diagram D. For more information, we refer to, for instance, Adámek et al. (1990, Theorem 12.3) or Leinster (2014, Proposition 5.1.26).  A category is called complete if it satisfies the equivalent properties in Proposition 5.37 and co-complete if Cop is complete. Observe that the dual of Proposition 5.37 says that a category has all small colimits if, and only if, all coequalizers and setsized coproducts exist, which happens if, and only if, all finite colimits and directed colimits exist. Note also that, if we specialize to the case of a thin category, then the equivalence of (i) and (ii) in Proposition 5.37 (and its dual version) reduce to a familiar characterization of complete lattices: A poset is a complete lattice if, and only if, all finite sets and all up-directed sets have a supremum. This fact was already stated in Exercise 1.2.12 and also plays an important role in the background in domain theory (Chapter 7). Exercises for Section 5.3 Exercise 5.3.1 Let f : L → M be a homomorphism between distributive lattices. (a) (b) (c)

(d) (e)

Prove that f is a monomorphism if, and only if, f is injective. Prove that if f is surjective, then f is an epimorphism. Consider an order-embedding e of the three-element chain 3 into the fourelement Boolean algebra, 2 × 2, which is clearly not surjective. Prove that e is an epimorphism in the category of distributive lattices. Conclude that in DL there exist morphisms which are both mono and epi, but not iso. Also, prove that e is a non-regular monomorphism in DL.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

186

5 Categorical Duality

Exercise 5.3.2 This exercise examines epimorphisms in categories of topological spaces; note in particular that the notion of “epimorphism” is not stable under taking full subcategories. (a) (b) (c)

Prove that in the category Top, any epimorphism is surjective. Prove that the epimorphisms in the category of Hausdorff topological spaces are exactly the continuous functions whose image is dense in the codomain. Deduce from the previous item that in Priestley, the epimorphisms are exactly the surjective continuous functions.

Exercise 5.3.3 and BA. (a)

(b) (c)

This exercise characterizes epimorphisms in the categories DL

Prove that in BA, the epimorphisms are exactly the surjective homomorphisms. Hint. One direction always holds. For the other, use Stone duality and the fact that the dual of a surjective homomorphism is injective. Let L be a distributive lattice and let e L : L → L − be its Boolean envelope (Definition 1.19). Prove that e L is an epimorphism in DL. Prove that in DL, a homomorphism f : L → M is an epimorphism if, and only if, the unique extension e M ◦ f : L − → M − of e M ◦ f : L → M − is surjective.

Exercise 5.3.4 Let f : A → B be a morphism in a category C. (a) (b) (c) (d)

Prove that if f is an isomorphism, then f is both a monomorphism and an epimorphism. Prove that if f has a left inverse (i.e., there exists g : B → A such that g f = 1 A), then f is a monomorphism. Formulate the dual statement of (b) and conclude that it is true. Consider the category with one object (i.e., a monoid), a countable set of morphisms { f n : n ∈ N}, composition given by f n ◦ f m := f n+m , and identity f 0 . Prove that, for every n ≥ 1, f n is mono and epi, but not iso.

Exercise 5.3.5 Prove that any equalizer is a monomorphism and any coequalizer is an epimorphism. Exercise 5.3.6 Suppose that q1 : A → Q1 and q2 : A → Q2 are both coequalizers for the same pair of maps p1, p2 : C → A. (a) (b)

Show that there exist morphisms α1 : Q1 → Q2 and α2 : Q2 → Q1 such that α1 q1 = q2 and α2 q2 = q1 . Prove that α2 α1 = idQ1 and α1 α2 = idQ2 . Hint. Show first that α2 α1 is a factorization of q1 through itself, that is, that α2 α1 q1 = q1 , and then use the uniqueness of α in the definition of coequalizers.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.3 Constructions inside Categories

187

Exercise 5.3.7 This exercise outlines a proof that the regular epimorphisms in Lat are exactly the surjective homomorphisms. To readers with more knowledge of universal algebra, it should be clear that the same argument applies in any category whose objects are the algebras in a variety and whose morphisms are the homomorphisms. (a) (b)

(c)

(d)

Prove that an equivalence relation ϑ on a lattice A is a congruence if, and only if, ϑ is a (bounded) sublattice of the product algebra A × A. Let p1, p2 : C → A be a pair of lattice homomorphisms. Define ϑ to be the smallest congruence on A which contains (p1 (c), p2 (c)) for every c ∈ C. Prove that the quotient q : A  A/ϑ is a coequalizer of (p1, p2 ). Conclude that regular epimorphisms are surjective. Let ϑ be a congruence on a lattice L. Prove that the quotient homomorphism q : A  A/ϑ is the coequalizer of the pair of maps p1, p2 : ϑ → A, where pi is the projection onto the i th coordinate. Conclude that a lattice homomorphism is a regular epimorphism if, and only if, it is surjective.

Exercise 5.3.8 Let C be a category and let I be a small category. (a)

(b) (c) (d)

Show that, for any object C ∈ C, the functor ΔC , defined by ΔC i = C for every object and ΔC x = 1C for every morphism, is a diagram, called the I-shaped constant diagram at C. Let D be a diagram. Show that a cone above D is essentially the same thing as a natural transformation p : ΔC ⇒ D for some object C of C. Reformulate the definition of limit purely in terms of natural transformations, without ever referring to specific objects and morphisms in I and C. Dualize this exercise for cones below and colimits.

Exercise 5.3.9 Prove that the Cartesian product of posets/lattices/distributive lattices is a product in the categorical sense of Definition 5.26. Exercise 5.3.10 Recall from Exercise 3.2.6 that the Cartesian product of a family of Priestley spaces, defined by equipping the product set with the product topology and the pointwise partial order, is again a Priestley space. Conclude that products exist in the category of Priestley spaces. Exercise 5.3.11 Let (Xi )i ∈I be a collection of Priestley spaces. For each i ∈ I, write L i for the lattice of clopen down-sets of Xi . Prove that the Priestley dual space   X of the product i ∈I L i is the coproduct i ∈I Xi in Priestley spaces. Show that there is an injective function from the disjoint union of the sets Xi to X, which is in general not surjective.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

188

5 Categorical Duality

Exercise 5.3.12 Let F : C → D be a functor and suppose that F has a right adjoint. Prove that F preserves any colimit that exists in C, that is, if D : I → C is a diagram in C and C is a colimit of D with a co-cone given by λ i : D(i) → C for each object i of I, then the image of this co-cone under F, F λ i : F D(i) → FC, gives a colimit of the diagram F ◦ D : I → D. Exercise 5.3.13 This exercise elaborates on the comment made in Example 4.6 to the effect that the two-element poset (2, ) appearing in the concrete incarnation of FDL (V ) derived in Section 4.1 is the poset dual to 3, the three-element lattice. (a) (b) (c)

Prove that 3 is a free distributive lattice over {p} by exhibiting a universal arrow.  Prove that for any set V we have FDL (V ) = v ∈V FDL ({v}). Conclude, via duality, that the Priestley dual of FDL (V ) is the product space (2, , δ)V , where δ is the discrete topology on 2, see Proposition 4.8.

Exercise 5.3.14 Let C be a category and let I be a small category. Denote by [I, C] the category of I-shaped diagrams in C, with natural transformations between them. (a)

(b) (c)

Show that there is a functor Δ : C → [I, C] which sends any object C of C to the constant diagram ΔC (see Exercise 5.3.8) and a morphism f : C → C  to the natural transformation ΔC → ΔC  that is f at every component. Suppose that Δ has a left adjoint, L : [I, C] → C. Prove that, for any diagram D ∈ [I, C], L(D) is a colimit of D. Similarly, prove that if R is a right adjoint for Δ, then R(D) is a limit of D, for any I-shaped diagram D in C.

Exercise 5.3.15 Let I be a preordered set that is directed, that is, any finite subset of I has an upper bound. Let π : I → J be the poset reflection of I (see Exercise 1.1.5). Let D : I → C be a diagram. (a) (b) (c)

Prove that J is directed. Prove that there is a diagram D  : J → C such that D  (π(i)) is isomorphic in C to D(i) for every i ∈ I. Suppose that D has a colimit in C. Prove that the same object gives a colimit of D  in C.

5.4 Priestley Duality Categorically In this section, we apply the categorical point of view introduced in this chapter to Priestley duality. While the general language of category theory is useful for expressing our duality theorem, we will also require some knowledge that seems to

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.4 Priestley Duality Categorically

189

be rather specific to Priestley duality. In particular, in this section, we further elaborate on what the correspondences between monomorphisms and epimorphisms on either side of the duality concretely look like in the case of Priestley duality. This yields connections between quotients and sublattices, and between subspaces and quotient lattices, which are crucial to some of the applications in Chapters 7 and 8 of this book. Further, we show how Priestley duality can be used to easily describe free distributive lattices. Priestley’s Dual 2-Equivalence Recall from Chapter 3 that, to any Priestley space X, we associated a distributive lattice ClpD(X ) of clopen down-sets of X and to any continuous order-preserving map f : X → Y , the distributive lattice homomorphism f −1 : ClpD(Y ) → ClpD(X ). In the reverse direction, if L is a distributive lattice, we associate to it the Priestley space X L (Definition 3.21), and if h : M → L is a distributive lattice homomorphism, then Proposition 3.27 showed that there is a unique continuous map f : X L → X M  = f −1 ( with the property that h(a) a ) for every a ∈ M. In this section, we will write X (h) for this unique map f . With the categorical language developed in this chapter, we can now properly state and prove Priestley’s duality theorem. Theorem 5.38 The assignments L → X L and (h : M → L) → (X (h) : X L → X M ) constitute a contravariant functor, X : DL → Priestley, and the pair (ClpD, X ) is a contravariant equivalence of categories. Proof To see that X is a functor, let k : N → M and h : M → L be morphisms in DL. Note that X (k) ◦ X (h) is a Priestley morphism for which it holds that, for any a ) = X (h) −1 (X (k) −1 ( a )) = X (h) −1 ( k (a)) = h(k (a)). a ∈ N, (X (k) ◦ X (h)) −1 ( By the uniqueness of X (h ◦ k), we must have X (h ◦ k) = X (k) ◦ X (h). The fact that X (1 L ) = 1XL for any distributive lattice L can be proved similarly. For any distributive lattice L, let α L : L → ClpD(X L ) be the homomorphism defined by a ∈ L →  a ∈ ClpD(X L ). For any Priestley space X, we recall the definition of an order-homeomorphism β X : X → XClpD(L) from Proposition 3.26; note that, compared to the notation of Definition 5.12, we here use the notation β for the inverse map. For any x ∈ X, β(x) is defined as the point associated to the prime filter B(x) = {K ∈ ClpD(L) | x ∈ K }, the clopen-down-set neighborhoods of x. To see that α : 1DL → ClpD ◦ X is a natural transformation, let h : M → L be a homomorphism between distributive lattices. Then, by definition of X, we have, for  showing the naturality of α. Finally, to see that a ) = h(a), any a ∈ M, that X (h) −1 ( β : 1Priestley → X ◦ ClpD is a natural transformation, let f : X → Y be a continuous order-preserving function. We want to show that the following diagram commutes:

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

190

5 Categorical Duality

X

f

Y

βX

XClpD(X)

βY X ( f −1 )

XClpD(Y ).

Let x ∈ X be arbitrary. Let us write F1 and F2 for the prime filters of ClpD(Y ) that correspond to the two points X ( f −1 )( β X (x)) and βY ( f (x)), respectively. We need to show that F1 = F2 . Recall from Exercise 3.2.11 that, for any homomorphism h : M → L, the prime filter corresponding to X (h)(x) is h−1 (Fx ). Thus, for any K ∈ ClpD(Y ), we have K ∈ F1 ⇐⇒ f −1 (K ) ∈ β X (x) ⇐⇒ x ∈ f −1 (K ) ⇐⇒ f (x) ∈ K which is equivalent to K ∈ F2 .



We now also prove that this duality has a further property, which, in categorical terms, makes it an equivalence enriched over partially ordered sets. The general definition of enriched equivalence is beyond the scope of this book; see, for example, Hofmann and Stubbe (2011) and Chikhladze et al. (2015) for more on the relationship between topological dualities and enriched categories, and the recent work Hofmann and Nora (2023) which develops Priestley duality in a quantale-enriched setting. In our categories DL and Priestley, the enriched structure amounts to the fact that the sets of morphisms are in fact posets. That is, if L and M are distributive lattices, then the set HomDL (L, M) is partially ordered by the pointwise order: For a parallel pair of homomorphisms h, h  : L → M, we say h ≤ h  if h(a) ≤ M h  (a) for every a ∈ L. Similarly, for any Priestley spaces X and Y , HomPriestley (X, Y ) is also ordered pointwise. The fact that Priestley duality is poset-enriched now boils down to the fact that the equivalence functors are anti-order-isomorphisms between the Hom-posets, as we show now. Proposition 5.39 For any distributive lattices L, M, the function X : HomDL (M, L) → HomPriestley (X L, X M ) is an anti-isomorphism of partial orders. That is, for any parallel pair of homomorphisms h, h  : M → L, we have h ≤ h  if, and only if, X (h  ) ≤ X (h), where the symbol ≤ denotes the pointwise order between lattice homomorphisms and between Priestley morphisms, respectively. We note that a similar result to this proposition, in the context of duality for relations and finite-meet-preserving functions, was already shown in the proof of Proposition 4.39; compare in particular equation (4.12) in that proof.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.4 Priestley Duality Categorically

191

Proof Recall (see Exercise 3.2.11) that X (h) sends any point x ∈ X L to the point X (h)(x) with associated prime filter FX (h)(x) := h−1 (Fx ). Suppose that h ≤ h  in the pointwise order and let x ∈ X L . For any a ∈ M, h(a) ∈ Fx implies h  (a) ∈ Fx , since Fx is an up-set. Hence, FX (h)(x) ⊆ FX (h )(x) , which means by definition of the order on Priestley spaces that X (h  )(x) ≤ X (h)(x). For the converse, suppose that h  h  in the pointwise order. Pick a ∈ M such that h(a)  h(a  ). By the distributive prime filter-ideal theorem (Theorem 3.10), pick a prime filter Fx in L such that h(a) ∈ Fx and h  (a)  Fx . For the corresponding point x ∈ X L , we have  h−1 (Fx )  (h  ) −1 (Fx ). Thus, X (h  )(x)  X (h)(x). We note in particular (see Exercise 5.4.1) that it follows from Proposition 5.39 that, for a pair of homomorphisms between distributive lattices f : L  M : g, ( f , g) is an adjoint pair if, and only if, the pair X ( f ) : X M  X L : X (g) is an adjoint pair between the underlying posets of the dual Priestley spaces, where X ( f ) is the lower adjoint and X (g) is the upper adjoint. Essentially this fact (phrased in the setting of spectral spaces) will be crucial in Section 7.3.

A Categorical Perspective on Subs and Quotients Theorem 5.40 Let L be a distributive lattice and X L its dual space. The complete lattice of sublattices of L is anti-isomorphic to the complete lattice of quotient spaces of X L and the complete lattice of quotients of L is anti-isomorphic to the complete lattice of closed subspaces of X L . Proof By Proposition 3.28, Priestley duality restricts to a dual equivalence between injective morphism into L and surjective morphisms from X L and also between surjective homomorphisms out of L and embeddings into X L . These lattices are concretely realized as stated in the theorem. To see that these are order-reversing bijections, see Proposition 4.26 and Proposition 4.15.  A Categorical Perspective on Operators We revisit the duality for unary operators described in Section 4.3, giving in particular the proof of the generalized duality Theorem 4.41, for which we now have the required terminology in place. You are invited to fill in some small details in Exercise 5.4.2. Proof of Theorem 4.41 Let us write DLm for the category of distributive lattices with finite-meet-preserving functions and PriestleyR↑ for the category of Priestley spaces with upward-compatible relations; the latter is indeed a category under relational composition, with the identity morphism on a Priestley space X given

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

192

5 Categorical Duality

by the upward Priestley compatible relation ≥ ⊆ X × X. We define a functor P : PriestleyR↑ → DLm by sending an object X to the lattice of clopen down-sets of X, and sending an upward Priestley compatible relation R ⊆ X ×Y to the finite-meetpreserving function P (R) := ∀R−1 [−] : ClpD(Y ) → ClpD(X ). It is straightforward to verify that P is a well-defined functor. Moreover, Proposition 4.39 shows that P is full and faithful. Finally, P is essentially surjective because any distributive lattice L is isomorphic (also in DLm ) to ClpD(X ) where X is its dual Priestley  space. It follows from Theorem 5.13 that P is part of a dual equivalence. We also show how a duality theorem for finite-join-preserving functions can be deduced from this in a purely abstract way. Theorem 5.41 The category of distributive lattices with finite-join-preserving functions is dually equivalent to the category of Priestley spaces with downward Priestley compatible relations. Proof The category DL j of distributive lattices with finite-join-preserving functions is isomorphic to the category of distributive lattices with finite-meet-preserving functions, via the isomorphism which sends a distributive lattice L to its opposite L op , and a morphism to itself. Similarly, the category of Priestley spaces with downward Priestley compatible relations is isomorphic to the category of Priestley spaces with upward Priestley compatible relations, via the isomorphism which sends a Priestley space X to the space with the same topology, and the opposite order. The claimed dual equivalence is obtained as the conjugate of the dual equivalence of Theorem 4.41 under these isomorphisms.  Since the proof of Theorem 5.41 is very abstract, we also give a more concrete description of the action of the dual equivalence functor, P : DL j → PriestleyR↓ , on morphisms. Fix two Priestley spaces X and Y . The bijection P (X, Y ) : PriestleyR↓ (X, Y ) → DL j (ClpD(Y ), ClpD(X )) may be defined as a composition of four bijections: id

PriestleyR↓ (X, Y ) −→ PriestleyR↑ (X op, Y op )

(5.1)

4.41 −→ DLm (ClpU(Y ), ClpU(X )) id

−→ DL j (ClpU(Y ) op, ClpU(X ) op ) c

−→ DL j (ClpD(Y ), ClpD(X )). Here, the first and third bijections are identity functions, using the obvious facts that a downward compatible relation from X to Y is the same as an upward

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.4 Priestley Duality Categorically

193

Priestley compatible relation from the opposite Priestley space X op to Y op , and that a finite-meet-preserving function from a lattice L to a lattice M is the same as a finite-join-preserving function from L op to M op . The second bijection is given by the dual equivalence of Theorem 4.41, modulo the fact that a clopen down-set of the Priestley space X op is the same thing as a clopen up-set of X, and similarly for Y . The fourth and last bijection is “conjugation by complementation:” c sends a finitejoin-preserving function h : ClpU(Y ) op → ClpU(X ) op to the finite-join-preserving function c(h) : ClpD(Y ) → ClpD(X ), defined by c(h)(K ) := X – h(Y – K ). Unraveling the definitions, the bijection P (X, Y ) can be seen to send R to ∃R−1 , as defined in (4.13) at the end of Section 4.3. Example 5.42 This example gives a functorial point of view on unary modal operators on a distributive lattice, which we studied in Section 4.3. Recall that a meet-semilattice is a poset (M, ≤) in which every finite set has an infimum; equivalently, a meet-semilattice is a commutative monoid (M, ∧, ) in which a∧a = a for all a ∈ M. Denote by SL the category of meet-semilattices with functions preserving all finite meets. For any meet-semilattice M, we make a set of symbols M = {a | a ∈ M } and define the distributive lattice F (M) := FDL (M)/ϑ M , where we recall from Section 4.1 that FDL (V ) denotes the free distributive lattice over a set V and we define ϑ M to be the congruence generated by the set of pairs of the form  

 G , {a | a ∈ G} , where G ranges over the finite subsets of M. Since this equivalence relation ϑ M exactly says that “box preserves finite meets,” the lattice F (M) can be thought of as the lattice generated by the act of “freely adding one layer of a box operation.” The function a → [a]ϑ M is an order embedding that preserves finite meets (see Exercise 5.4.3). The constructor F may be extended to a functor from the category of meetsemilattices to the category of distributive lattices, in such a way that, for any finitemeet-preserving function h : M → M  of meet-semilattices, we have F h([a]ϑ M ) = [h(a)]ϑ M  (see Exercise 5.4.3). Pre-composing F with the forgetful functor from distributive lattices to meet-semilattices, we obtain a functor T from DL to itself. This functor T is closely related to the category of positive -modal algebras, where a positive -modal algebra is defined as a pair (L, ), with L a distributive lattice and  : L → L a finite-meet-preserving function. Indeed, there is a one-toone correspondence between  operations on a distributive lattice L and so-called T -algebra structures based on L (see Exercise 5.4.3).

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

194

5 Categorical Duality

The functor T : DL → DL is one of the program constructors in Abramsky’s program logic, which we will study in Chapter 7. It is dual to the Vietoris space construction of topology, and is known in domain theory as the Smyth powerdomain, see also Definition 6.21, Chapter 6. For a survey of dualities for modal logics in a general categorical setting, and the related development of coalgebraic modal logic, see Venema (2007). Exercises for Section 5.4 Exercise 5.4.1 Prove that, for any pair of homomorphisms between distributive lattices f : L  M : g, ( f , g) is an adjoint pair if, and only if, the pair X ( f ) : X M  X L : X (g) is an adjoint pair between the underlying posets of the dual Priestley spaces, where X ( f ) is the lower adjoint and X (g) is the upper adjoint. Hint. Combine Proposition 5.39 and Exercise 1.1.8 of Chapter 1. Exercise 5.4.2 Theorem 4.41. (a) (b)

This exercise asks you to fill in a few details of the proof of

Prove that PriestleyR↑ , the category of Priestley spaces with upward Priestley compatible relations, is indeed a category. Prove that P as defined in the proof of Theorem 4.41 is indeed a well-defined functor.

Exercise 5.4.3 For a meet-semilattice M, consider the lattice F (M) defined in Example 5.42. For every a ∈ M, define e(m) := [m]ϑ M. (a)

(b) (c)

Prove that, for any distributive lattice L and any finite-meet-preserving function f : M → L, there exists a unique DL homomorphism f¯ : F (M) → L such that f¯ ◦ e = f . Prove that e is finite-meet-preserving, injective, and that the image of e generates F (M). Prove that the object assignment F extends to a functor SL → DL, which is left adjoint to the forgetful functor U : DL → SL that sends any distributive lattice (L, , ⊥, ∧, ∨) to its meet-semilattice reduct, (L, , ∧), and any lattice homomorphism to itself, now viewed as a meet-semilattice homomorphism. Note. It is a general fact in category theory that “universal arrows yield left adjoints,” see, for example, Borceux (1994, Section 3.1) for the general theory, of which you only prove a specific instance here.

For a functor 5 T from C to itself, an algebra for the functor T, or simply T-algebra, is defined as a pair ( A, h), where A is an object in C and h : T ( A) → A is a morphism 5

There is a related, but different, notion of algebra for a monad, which we do not treat in this book (at least not explicitly).

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

5.4 Priestley Duality Categorically

195

in C. A morphism of T-algebras, g : ( A, h) → ( A, h  ), is defined to be a morphism g : A → A in C so that g ◦ h = h  ◦ T (g); that is, the following diagram commutes: T ( A)

T (g)

T ( A ) h

h A

g

A .

Recall that a positive -modal algebra is a pair ( A, ) where A is a distributive lattice and  : A → A preserves finite meets. Write T for the composite functor U

F

DL → SL → DL. The rest of this exercise makes precise the idea that T freely adds one layer of the unary modal  operator. For further details, see Abramsky (1988), Ghilardi (1995), and Bezhanishvili and Kurz (2007). The analogous construction for a binary implication-type operator is discussed and applied in Section 6.4. (d) Show that the category of T -algebras is isomorphic to the category of positive -modal algebras. We will now construct the free positive -modal algebra over a distributive lattice by iterated application of the functor F := 1DL +T . Here, 1DL is the identity functor and + is the coproduct computed in DL; more explictly, for any distributive lattice L, F (L) is the lattice L + T (L), and if h : L → M is a lattice homomorphism, then F (h) : L + T (L) → M + T (M) is given by h + T (h). Let L be a distributive lattice, and consider the countable chain of distributive lattices e0 e1   e2 L → L + T (L) → L + T L + T (L) → · · · , where the first embedding e0 is the inclusion of L in the coproduct L + (−) and the subsequent embeddings are defined by en+1 := e0 + T (en ). (e) Show that the colimit of the chain is the free positive -modal algebra over L. (f) Show that, when viewing the chain as a tower of sublattices of the free positive -modal algebra over L, for each n ≥ 0, the nth lattice consists precisely of the elements that can be described by a term in which the maximum nesting depth of the operation  is less than or equal to n.

https://doi.org/10.1017/9781009349680.006 Published online by Cambridge University Press

6 Omega-Point Duality

In this chapter, we consider a Stone-type duality that applies to spaces more general than the ones dual to distributive lattices. We begin (Section 6.1) by giving Stone’s original duality (Stone, 1937b) for bounded distributive lattices with a class of (unordered) topological spaces that we call spectral spaces in this book. In particular, we show in Theorem 6.4 that the category of spectral spaces is isomorphic to the category of Priestley spaces, establishing exactly how Stone and Priestley duality for bounded distributive lattices are two presentations of the same mathematical result. This recasting of Priestley duality uses the correspondence between compact ordered spaces and stably compact spaces already discussed in Section 2.3. Then, we introduce an adjoint pair of functors (Section 6.2) between the category of topological spaces and a certain category of complete lattices and we show that this adjunction restricts to a duality between so-called sober spaces and spatial frames (Section 6.3). Throughout these sections, we in particular pay attention to how all these dualities fit together. In the final section, Section 6.4, as an application of the duality between spectral spaces and distributive lattices, we study two classical constructions on spectral spaces: powerdomains and function spaces. This leads us to introduce the notion of preserving joins at primes, which will play a role in both of the application chapters, Chapters 7 and 8. 6.1 Spectral Spaces and Stone Duality In the Priestley duality of Chapter 3, we obtain certain compact ordered spaces as duals of distributive lattices. However, in the original formulation by Stone, the dual spaces of distributive lattices are certain stably compact spaces (Definition 2.18). In fact, the relationship between Stone’s dual spaces and Priestley spaces is given by restricting the bijective correspondence between compact ordered spaces and stably compact spaces established in Theorem 2.21 of Chapter 2. The precise relationship between distributive lattices, Priestley spaces, and spectral spaces will be given in the commutative triangle of equivalences in Figure 6.1. The main

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.1 Spectral Spaces and Stone Duality

197

technical ingredient needed for establishing these relationships is the following. Recall from Section 2.3 that, for any stably compact space (X, τ), we write τ ∂ for the co-compact dual topology of τ and τ p for the patch topology τ ∨ τ ∂ . Theorem 6.1 Let (X, τ) be a stably compact space. The following statements are equivalent: (i) (ii) (iii) (iv)

(X, τ p, ≤τ ) is a Priestley space; (X, τ p, ≥τ ) is a Priestley space; (X, τ) has a base of compact-opens; and (X, τ ∂ ) has a base of compact-opens.

Proof Since the definition of a Priestley space is self-dual with respect to orderduality, it follows that (i) and (ii) are equivalent. Since, under the correspondence of Theorem 2.21, (X, τ) is the stably compact space corresponding to (X, τ p, ≤τ ) and (X, τ ∂ ) is the stably compact space corresponding to (X, τ p, ≥τ ), it suffices to show that (iii) is equivalent to (i) and/or (ii) in order to show that all four statements are equivalent. So, suppose (X, τ) is a stably compact space for which (X, τ p, ≤τ ) is a Priestley space and let x ∈ X and U ∈ τ with x ∈ U. We will construct a compact-open set V such that x ∈ V ⊆ U. Note that C := U c is a closed down-set which does not contain x. Hence, for each y ∈ C we have x  y and thus there is a clopen up-set Vy of the Priestley space (X, τ p, ≤τ ) with x ∈ Vy and y ∈ Vyc . It follows that {Vyc }y ∈ C is an open cover of C in (X, τ p, ≤τ ). Since C is closed and thus compact, it follows that there is a finite subset F ⊆ C so that {Vyc }y ∈F is a cover of C. Consequently, the  set V := {Vy | y ∈ F} is a clopen up-set of (X, τ p, ≤τ ) with x ∈ V ⊆ U. Finally, since V is an open up-set in (X, τ p, ≤τ ), it is open in the topology (τ p ) ↑ , which is equal to τ by Proposition 2.20. Also, since V is closed in (X, τ p, ≤τ ), it is compact in (X, τ p, ≤τ ), and thus also in the smaller topology of (X, τ), since the compact topologies are a down-set in Top(X ) (see Exercise 2.2.1). We have thus shown that (X, τ) possesses a base of compact-opens. For the converse, suppose (X, τ) is a stably compact space with a base of compactopens, and let x, y ∈ X with x τ y. By definition of the specialization order, it follows that there is an open U ∈ τ with x ∈ U and y  U. Also, since (X, τ) has a base of compact-opens, it follows that there is a compact-open V ⊆ X with x ∈ V ⊆ U. The fact that V is open in (X, τ) implies that it is an open up-set in (X, τ p, ≤τ ). Furthermore, since V is compact and open, and thus in particular saturated, it is closed in (X, τ ∂ ), and thus also in (X, τ p, ≤τ ). Altogether, we have that V is a clopen up-set in (X, τ p, ≤τ ) and that x ∈ V and y  V , as required.  We recall from Chapter 3 that the morphisms of Priestley spaces are the continuous and order-preserving maps, see also Example 5.2. Thus, in order to obtain a (non-full) subcategory of Top that is dual to the category of bounded distributive

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

198

6 Omega-Point Duality

lattices, we have to consider the corresponding stably compact spaces with proper maps (see Exercise 2.3.15). Finally, if X and Y are stably compact spaces satisfying the equivalent statements of Theorem 6.1, one can simplify the definition of proper maps: a map f : X → Y is a proper continuous map if, and only if, the inverse image of a compact-open is compact-open (see Exercise 6.1.1). In this context, proper maps are also known as spectral maps. The above considerations naturally suggest the following definition of the category of spectral spaces. A shorter but more ad hoc definition can be found in Exercise 6.1.3; alternative definitions in the literature often also explicitly include the statement that a spectral space is sober (see Definition 6.18), but we will derive this as a consequence of our definition (see Exercise 6.3.5). The spaces we call spectral spaces are also known as coherent spaces or (non-Hausdorff) Stone spaces in less recent literature, but these terms have also been used with other meanings, so we will avoid using them. Definition 6.2 A spectral space is a stably compact space satisfying the equivalent statements of Theorem 6.1. A spectral map is a continuous function between spectral spaces such that the inverse image of any compact-open set is compact (and open). We denote by Spec the category of spectral spaces with spectral maps. Definition 6.3 Let L be a distributive lattice. The Stone dual space of L or spectral space of L, denoted St(L), is a topological space (X, τ), where X comes with bijections F( ) : X → PrFilt(L), I( ) : X → PrIdl(L), h ( ) : X → HomDL (L, 2) so that for all x ∈ X and all a ∈ L, we have a ∈ Fx

⇐⇒

a  Ix

⇐⇒

h x (a) = 

and the topology τ is generated by the sets η(a) = {x ∈ X | a ∈ Fx } = {x ∈ X | a  Ix } = {x ∈ X | h x (a) = }, where a ranges over the elements of L. This object assignment extends to a contravariant functor St : DL → Spec, where a homomorphism h : L → M is sent to the function f : X M → X L given by requiring that Ff (x) = h−1 (Fx ), for every x ∈ X M . Given a topological space X we denote by K O(X ) the collection of compact-open subsets of X. Note that, if (X, τ) is a spectral space with specialization order ≤τ , then K O(X ) ordered by inclusion is the distributive lattice which corresponds under Priestley duality to the Priestley space (X, τ p, ≥τ ). Indeed, the clopen down-sets of (X, τ p, ≥τ ) are the clopen up-sets of (X, τ p, ≤τ ), which are exactly the compactopen sets of (X, τ) by Proposition 2.22(c). By definition, any proper continuous map

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.1 Spectral Spaces and Stone Duality

199

f : X → Y yields a distributive lattice homomorphism f −1 : K O(Y ) → K O(X ). In summary, we have the following commutative triangle of categorical (dual) equivalences, where the equivalence between Priestley and Spec is in fact an isomorphism, see Theorem 6.4. Priestley Pr (op ClpD DL

p ( ) ↓  ( ) , ≥( ) 

St (op KO Spec Figure 6.1 A commutative triangle of equivalences. The equivalence between Priestley and Spec is an isomorphism of categories, see Theorem 6.4. The dual equivalence between DL and Priestley is the one described in Theorem 5.38. The functors between DL and Spec have been defined in this section; it is a dual equivalence because it is the composition of a dual equivalence with an isomorphism.

Theorem 6.4 The categories Spec and Priestley are isomorphic. On objects, the isomorphisms are given, respectively, by the restriction of the patch construction for stably compact spaces, equipped with the reverse of the specialization order: ( ) p, ≥( )  : Spec → Priestley (X, τ) → (X, τ ∨ τ ∂, ≥τ ) and by intersecting with the dual Alexandrov topology of the order: ( ) ↓ : Priestley → Spec (X, ρ, ≤) → (X, ρ ∩ D(X, ≤)). On morphisms, the isomorphism of categories is simply the identity. Proof By Theorem 6.1 and the definition of spectral spaces, it follows that all spectral spaces are stably compact spaces and that the correspondence of Theorem 2.21, applied with a reversal of the orderings, between stably compact spaces and compact ordered spaces provides a bijective correspondence between spectral spaces and Priestley spaces. Furthermore, Exercise 2.3.15 shows that this bijective correspondence extends to an isomorphism of categories between the category

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

200

6 Omega-Point Duality

Priestley and that of spectral spaces with proper maps. Finally, by Exercise 6.1.1, proper maps between spectral spaces are precisely those for which the inverse image of a compact-open is compact-open, which in turn, by Definition 6.2, are the morphisms of the category Spec.  Exercises for Section 6.1 Exercise 6.1.1 Show that a map between spectral spaces is proper if, and only if, the inverse image of a compact-open is compact-open. Further show that a map between Boolean spaces is proper if, and only if, it is continuous. Exercise 6.1.2 Show, by giving an example, that there may exist continuous maps f : X → Y between spectral spaces X and Y which are not proper. Exercise 6.1.3 Show that a topological space X is a spectral space if, and only if, it is well filtered and K O(X ) is a bounded sublattice of P (X ) and a base for the topology of X. Note. We will derive an alternative characterization of spectral spaces, as a subclass of T0 sober spaces, in Exercise 6.3.5 below. Exercise 6.1.4 Using Priestley duality, Theorem 2.21 of Chapter 2, and Theorem 6.1, show that: (a)

The assignment X → K O(X ) may be extended to a contravariant functor from Spec to DL by sending a proper map f : X → Y to the map K O( f ) : K O(Y ) → K O(X ), U → f −1 (U).

(b)

The pair of functors St and K O yield a duality between the categories Spec and DL.

Exercise 6.1.5 Show that if (X, τ) is a spectral space, then the co-compact dual of τ is generated by the complements of the compact-opens of (X, τ). That is, τ ∂ = U c | U ∈ K O(X, τ). 6.2 The Omega-Point Adjunction The collection of compact-open subsets of a spectral space is closed under finite intersections and finite unions and is thus a distributive lattice. In fact, Stone duality tells us, among other things, that any distributive lattice occurs as K O(X ) for an, up to homeomorphism, unique spectral space X. Topological spaces that are not spectral spaces will in general not have a base of compact-opens, but they will have lots of bases that are sublattices of the power set. Using these bases, equipped with certain proximity relations, leads to an alternative duality for stably compact spaces (Smyth, 1992; Jung and Sünderhauf, 1996).

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.2 The Omega-Point Adjunction

201

If a compact space has a base of compact-opens, then, by compactness, this is necessarily the smallest base which is closed under finite unions. Topological spaces that are not spectral spaces will in general also not have a smallest base. To obtain a canonical base beyond spectral spaces, we are forced to take the base of all opens. But the collection of open subsets of a topological space is closed under arbitrary unions and is thus a complete lattice in the inclusion order. As a consequence, if we want to extend Stone duality to a duality between certain lattices and spaces beyond those with bases of compact-opens, we are forced to consider a category of complete lattices. Definition 6.5 A frame is a complete lattice in which finite meets distribute over arbitrary joins, that is, it satisfies the JID law (see (1.4)). A map between frames is said to be a frame homomorphism provided it preserves finite meets and arbitrary joins. We denote by Frame the category of frames with frame homomorphisms. Example 6.6 Given a topological space X, the collection Ω(X ) of all opens is closed under finite intersections and arbitrary unions. Since, in the power set lattice P (X ), (arbitrary) intersections distribute over arbitrary unions, Ω(X ) is a frame. Also, by the very definition of continuity, a continuous map f : X → Y yields a map Ω( f ) : Ω(Y ) → Ω(X ) given by U → f −1 [U]. Since the inverse image map preserves (arbitrary) intersections and arbitrary unions, the map Ω( f ) is a frame homomorphism. Proposition 6.7 Ω : Top → Frame is a contravariant functor. Proof It is straightforward to verify that Ω(idX ) = idΩ(X) for any topological space X and that Ω( f ◦ g) = Ω(g) ◦ Ω( f ) whenever g : X → Y and f : Y → Z are continuous maps.  If we want a way to recapture a space from its frame of open sets, first, we need a way of distinguishing points using open sets. As soon as a space is T0 this is possible via neighborhood filters. Note that a similar definition, but restricted to clopen down-sets, was already used when we proved Priestley duality (Proposition 3.26). Definition 6.8 Let X be a topological space and x ∈ X. The neighborhood filter of x is given by N (x) = {U ∈ Ω(X ) | x ∈ U}.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

202

6 Omega-Point Duality

Definition 6.9 Let L be a frame. A subset F ⊆ L is a completely prime filter of L if it is a filter satisfying, for all S ⊆ L,  S ∈ F, then S ∩ F  ∅. if Note that a completely prime filter is necessarily proper, since ⊥ is the supremum of the empty set, and can therefore never be an element of a completely prime filter. We denote by CompPrFilt(L) the set of completely prime filters of L. The proof of the following proposition is left as Exercise 6.2.3. Proposition 6.10 Let X be a topological space and x ∈ X. Then, N (x) is a completely prime filter of Ω(X ). Furthermore, X is T0 if, and only if, the assignment x → N (x) is injective. Recall from Chapter 1 that, for any distributive lattice L, M(L) denotes the set of (finitely) meet-prime elements of L. Proposition 6.11 equivalent:

Let L be a frame and F ⊆ L. The following statements are

the set F is a completely prime filter of L; the complement L – F of F is a principal down-set ↓m, for some m ∈ M(L); and (iii) the characteristic function χ F : L → 2, where χ F (a) = 1 if, and only if, a ∈ F, is a frame homomorphism.  Proof First suppose F is a completely prime filter of L. Define m := (L – F). Clearly, L – F ⊆ ↓m. Since F is completely prime and (L – F) ∩ F = ∅, it follows that m  F. Therefore, since L – F is a down-set containing m, we have ↓m ⊆ L – F, and thus ↓m = L – F. Finally, to show that m is meet prime, let M ⊆ L be finite and suppose that m   m for all m  ∈ M. Then, M ⊆ L − ↓m = F, so, since F is a   filter, M ∈ F, that is, M  m. Thus, m ∈ M(L), and we have shown that (i) implies (ii). Now, suppose that L – F = ↓m for some m ∈ M(L). To see that χ F preserves ar  bitrary joins, note that, for any A ⊆ L, χ F ( A) = 1 if, and only if, A  m, if, and  only if, a  m for some a ∈ A, which is equivalent to χ F [A] = 1. That χ F preserves finite meets follows directly from the definition of meet prime: for any finite   A ⊆ L, χ F ( A) = 1 if, and only if, A  m, which, since m ∈ M(L), happens  if, and only if, a  m for every a ∈ A, and this is in turn equivalent to χ F [A] = 1. For the last implication, suppose that χ F : L → 2 is a frame homomorphism. Since χ F preserves finite meets, it follows that F = χ−1 F (1) is a filter. If A ⊆ L   with A ∈ F, then χ F ( A) = 1 and thus χ F (a) = 1 for some a ∈ A. That is, A ∩ F  ∅ and thus F is completely prime.  (i) (ii)

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.2 The Omega-Point Adjunction

203

We are now ready to give the definition of the Pt functor which takes us from frames back to topological spaces. By Proposition 6.11, there are bijections between the set of homomorphisms to the frame 2, the set of completely prime filters, and the set of meet primes. The space of points of L can be based on any of these three sets. As we did for the Priestley dual space in Chapter 3, we give a “neutral” description here, leaving it as an exercise (Exercise 6.2.4) to check the details of the equivalence of the descriptions, since the proofs are very similar to those already given in Section 3.2 for Priestley duality. Note, however, that we only obtain a contravariant adjunction here; we will show how it restricts to a duality in Section 6.3. Definition 6.12 Given a frame L, we denote by Pt(L) a topological space which is determined up to homeomorphism by the fact that it comes with three bijections F : Pt(L) → CompPrFilt(L), m : Pt(L) → M(L), and f (−) : Pt(L) → HomFrame (L, 2) satisfying for any x ∈ Pt(L) and a ∈ L, a ∈ F (x) ⇐⇒ a  m(x) ⇐⇒ f x (a) = 1, and the opens of Pt(L) are the sets  a for a ∈ L, defined, for x ∈ Pt(X ), by def

x∈ a ⇐⇒ a ∈ F (x) ⇐⇒ a  m(x) ⇐⇒ f x (a) = 1. The assignment L → Pt(L) extends to a contravariant functor Pt : Frame → Top by sending a frame homomorphism h : L → M to the continuous map Pt(h) : Pt(M) → Pt(L) given by Pt(h)(x) = y

⇐⇒

h−1 (F (y)) = F (x)

⇐⇒

h−1 (↓m(y)) = ↓m(x)

⇐⇒

f y = f x ◦ h.

Theorem 6.13 The functors Pt : Frame  Top : Ω form a contravariant adjunction with corresponding natural transformations given, for L a frame, by η L : L → Ω(Pt(L)) a → {x ∈ Pt(L) | a ∈ F (x)} and, for X a topological space, ε X : X → Pt(Ω(X )) sends a point x of X to the element of Pt(Ω(X )) which corresponds to the neighborhood filter of x. Proof In order to check that the pair (Pt, Ω) forms a contravariant adjunction it suffices to check that:

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

204

(a)

6 Omega-Point Duality

The assignments η and  define natural transformations, that is, for each frame homomorphism h : L → M and each continuous map f : X → Y , the following diagrams commute: L

ηL

Ω(Pt(L)) Ω(Pt(h))

h M (b)

X

Pt(Ω(X ))

f

Ω(Pt(M))

ηM

εX

Pt(Ω( f )) Y

εY

Pt(Ω(Y )).

The triangle identities hold, that is, for each frame L and each space X, the following diagrams commute: Pt(L)

ε Pt(L)

1

Pt(Ω(Pt(L)))

Ω(X )

Pt(η L )

η Ω(X)

Ω(Pt(Ω(X ))) Ω(ε X )

1

Pt(L)

Ω(X ).

We leave these verifications as an exercise for the reader.



Exercises for Section 6.2 Exercise 6.2.1 Give an example of a topological space X such that: (a) (b) (c)

Ω(X ) is not closed under arbitrary intersections; Ω(X ) does not satisfy the Meet Infinite Distributive law (that is, the order dual of JID); and there is a frame homomorphism from Ω(X ) into 2 which does not preserve arbitrary meets.

Exercise 6.2.2 Let X be a topological space and let L = Ω(X ). Prove that, for any    S ⊆ L, the infimum of S is the interior of the set U ∈S U, that is, S = int( S). Exercise 6.2.3 Prove Proposition 6.10. Exercise 6.2.4 Show that Pt(L) is well defined and that it is a topological space. Also show that the bijection f ( ) : Pt(L) → HomFrame (L, 2) embeds Pt(L) as a closed subspace of the space S L , where S denotes the Sierpinski space (see Example 2.19). Further, show that, for any frame homomorphism h, Pt(h) is well defined and that it is a continuous map. Finally, show that Pt is a contravariant functor as stated in Definition 6.12. Exercise 6.2.5 Complete the proof of Theorem 6.13.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.3 The Omega-Point Duality

205

6.3 The Omega-Point Duality Recall from Corollary 5.19 that any contravariant adjunction, such as the one given by the functors Pt and Ω in Section 6.2, restricts to a maximal duality. This associated duality is obtained by restricting the functors to the full subcategories given by the objects for which the component of the natural transformation from the identity to the composition of the functors in the appropriate order is an isomorphism. Accordingly, we want to characterize those frames L for which η L : L → Ω(Pt(L)) is an isomorphism and those topological spaces X for which ε X : X → Pt(Ω(X )) is a homeomorphism. Proposition 6.14 Let L be a frame. The following statements are equivalent: (i) (ii) (iii) (iv)

η L : L → Ω(Pt(L)) is an isomorphism; η L is injective; for all a, b ∈ L, if b  a, then there is x ∈ Pt(L) with b ∈ F (x) and a  F (x); and  for all a ∈ L, a = (M(L) ∩ ↑a).

Proof By definition of Pt(L), the opens of this space are precisely the sets in the image of η L and thus η L is always surjective. It follows that items (i) and (ii) are equivalent since a frame homomorphism is an isomorphism if, and only if, it is both injective and surjective. Since η L is always a frame homomorphism, if it is injective, then it is an order embedding (see Exercise 1.2.7). On the other hand, if it is an order embedding, then it is clearly injective. So (ii) is equivalent to saying that η L is an order embedding. Further, again because η L is always a frame homomorphism, it is always order preserving. So (ii) is equivalent to η L being order reflecting. But (iii) is exactly the contrapositive of the statement that η L is order reflecting. We now use Proposition 6.11 to prove the equivalence of (iii) and (iv). In order to prove that (iii) implies (iv), we reason by contraposition: Suppose that (iv) fails, we show that (iii) also fails. Let a ∈ L be such that (iv) fails. Denote  by b the element (M(L) ∩ ↑a). We must then have b  a, since a ≤ b always holds. Now, let x ∈ Pt(L) be arbitrary such that b ∈ F (x). To show that (iii) fails, we need to show that a ∈ F (x). By Proposition 6.11, pick m ∈ M(L) such that F (x) = (↓m) c . Then, since b ∈ F (x), we have b  m, so in particular a  m, by the very definition of b. Thus, a ∈ F (x) as required. For the converse, suppose (iv) holds and let a, b ∈ L with b  a. Since a =   (M(L) ∩ ↑a), we have b  (M(L) ∩ ↑a) and thus there is m ∈ M(L) with a ≤ m but b  m. Again by Proposition 6.11, F = (↓m) c is a completely prime filter of L, and thus there is an x ∈ Pt(L) with F (x) = F. It follows that b ∈ F (x) and a  F (x) and thus (iii) holds. 

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

206

6 Omega-Point Duality

Remark 6.15 As we saw in Section 1.3, in a finite distributive lattice, each element is a finite join of join-irreducible elements and a finite meet of meet-irreducible elements. This is not the case for frames in general (see Exercise 6.3.1). Definition 6.16 A frame is said to be spatial provided it satisfies the equivalent conditions of Proposition 6.14. We denote by SpFrame the full subcategory of Frame given by the spatial frames. As we have already observed, the natural morphism ε X : X → Pt(Ω(X )) is injective if, and only if, X is a T0 space, see Proposition 6.10. Proposition 6.17 Let X be a T0 space. The following statements are equivalent: (i) (ii) (iii) (iv)

ε X : X → Pt(Ω(X )) is a homeomorphism; ε X is surjective; for all y ∈ Pt(Ω(X )) there exists an x ∈ X so that F (y) = N (x); and the join-irreducible elements of the lattice C(X ) of closed sets in X are precisely the closures of points, {x} = ↓x, for x ∈ X.

Proof Recall that the opens of Pt(Ω(X )) are, by definition, the sets of the form  = {y ∈ Pt(L) | U ∈ F (y)}, where U ∈ Ω(X ). We have, for any x ∈ X and U U ∈ Ω(X ),  ⇐⇒ U ∈ F (ε X (x)) = N (x) ε X (x) ∈ U ⇐⇒ x ∈ U. Thus, ε X is always an open continuous embedding of topological spaces and therefore it is clear that (i) is equivalent to (ii). Also, by definition of ε X , a point x ∈ X is sent to the unique point y ∈ Pt(Ω(X )) with the property that F (y) = N (x). Thus, (iii) is simply spelling out the statement that ε X is surjective and therefore (ii) and (iii) are equivalent. To prove that (iii) implies (iv), suppose that (iii) holds and let C ∈ C(X ) be join irreducible. Then, since the lattice of closed sets of X is order dual to the lattice of open sets of X, we have that U = C c is a meet-irreducible element of the frame Ω(X ). By Proposition 6.11, it follows that the set FC = (↓U) c = {V ∈ Ω(X ) | V  U} = {V ∈ Ω(X ) | C ∩ V  ∅} is a completely prime filter of Ω(X ). Now, let y ∈ Pt(Ω(X )) be the point of Ω(X ) corresponding to FC , that is, the point with F (y) = FC . Since (iii) holds, there is x ∈ X with FC = N (x). That is, for all V ∈ Ω(X ) we have C ∩V  ∅

⇐⇒

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

x ∈ V.

6.3 The Omega-Point Duality

207

We use this to show that C = {x}. To this end, let D be any closed subset of X and let V = D c be the complementary open set. Then, we have the following string of equivalences C ⊆ D ⇐⇒ C ∩ V = ∅ ⇐⇒ x  V ⇐⇒ x ∈ D. Thus, we obtain the desired conclusion C= {D ∈ C(X ) | C ⊆ D} = {D ∈ C(X ) | x ∈ D} = {x} and every join-irreducible closed set is the closure of a singleton. The fact that closures of singletons always are join irreducible in the lattice of closed sets is left for the reader as Exercise 6.3.4. To prove that (iv) implies (iii), let y ∈ Pt(Ω(X )) and let F = F (y) be the corresponding completely prime filter of Ω(X ). By Proposition 6.11, the complement of F is the down-set of some open set U which is meet irreducible in the lattice Ω(X ). It follows that C = U c is a closed subset of X which is join irreducible in the lattice C(X ) of closed sets in X. By (iv), there exists x ∈ X with C = {x}. Now, let V ∈ Ω(X ). Then we have V  F ⇐⇒ V ⊆ U ⇐⇒ C ∩ V = ∅ ⇐⇒ {x} ∩ V = ∅ ⇐⇒ x  V and thus V ∈ F if, and only if, x ∈ V so that F = N (x) as required.



Definition 6.18 A topological space is said to be sober provided it is a T0 space and satisfies the equivalent conditions of Proposition 6.17. We denote by Sober the full subcategory of Top given by the sober topological spaces. It now follows, using Corollary 5.19, that the functors Ω and Pt, properly restricted and co-restricted, yield a duality between the category of spatial frames and the category of sober spaces. Theorem 6.19 The Ω-Pt adjunction cuts down to a duality between the category of spatial frames with frame homomorphisms and the category of sober topological spaces with continuous maps. Ω SpFrameop

Sober. Pt

We end this section with a summary of the relationship between the Ω-Pt, the Stone, and the Priestley dualities. Recall that we already showed how Stone and Priestley duality for distributive lattices relate in Figure 6.1. We now add the Ω-Pt duality into the mix. Since the Ω-Pt duality is for unordered topological spaces rather than ordered spaces, we compare it to Stone duality rather than to Priestley duality. As we ask you

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

208

6 Omega-Point Duality

to show in Exercise 6.3.5, every spectral space is sober, thus Spec is a subcategory of Sober. Note that it is not a full subcategory, since not every continuous map between spectral spaces is proper (see Exercise 6.1.2). We will denote by Specc the full subcategory of Top based on spectral spaces, that is, the category whose objects are spectral spaces and whose morphisms are all continuous functions. Exercise 6.3.7 outlines a generalization of Stone duality for the category Specc – the dual category of distributive lattices then has certain relations as its morphisms. The category Specc can be used to give a categorical description of the category of stably compact spaces with continuous functions: It is the so-called Karoubi envelope or splitting by idempotents of Specc . The topological fact underlying this theorem is that every stably compact space is a continuous retract of a spectral space, see for example Johnstone (1986, Theorem VII.4.6) and van Gool (2014, Section 2.2) for more information. The Ω-Pt duality is not directly a generalization of Stone duality as, under the Ω-Pt duality, a spectral space is sent to its entire open set lattice rather than just to the lattice of compact-opens. However, the category DL embeds into the category Frame via the ideal completion A → Idl( A), which freely adds directed joins to A. A distributive lattice can be recovered from its ideal completion, as follows. We call an element k of a frame F compact provided that, for every directed subset S of F, we  have k ≤ S implies k ≤ s for some s ∈ S, also see Section 7.1 for more on compact elements in the more general setting of directedly complete posets. In particular, we will encounter in Chapter 7 a more general version of this construction, P → Idl(P), which turns an arbitrary poset into a so-called algebraic domain. For an arbitrary frame F, we denote by K(F) the set of compact elements of F, which always forms a join-subsemilattice of F. For any distributive lattice A, the compact elements in Idl(A) are the principal ideals, thus we have A  K(Idl( A)). Finally, calling coherent frames 1 those frames whose compact elements form a sublattice which generates the frame by directed joins, we obtain Figure 6.2, which illustrates how to move back and forth between the Stone duality and the Ω-Pt duality on objects. K O-St

DL

Sober ⊇ Spec

Idl-K Ω-Pt

CohFr ⊆

SpFrame

Figure 6.2 Comparing the Ω-Pt and Stone duality on objects.

This diagram is not the whole story as it does not specify what happens with morphisms. Stone duality acts on maps in Spec, which is not a full subcategory 1

Coherent frames are also known as arithmetic frames in the literature.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.3 The Omega-Point Duality

209

of the category Sober of sober spaces with continuous maps. The Ω-Pt duality, on the other hand, works on the full subcategory Specc of Sober on the spectral spaces. To get the morphisms between coherent frames that are dual to the proper maps, we need to restrict to the subcategory of CohFr with the same objects, but in which the morphisms are required to send compact elements to compact elements; such morphisms are called coherent morphisms. An alternative way of seeing this restriction of the duality between Specc and CohFr to a duality for Spec is bitopological: Coherent morphisms between coherent frames correspond precisely to those continuous maps between spectral spaces that are also continuous with respect to the co-compact dual spectral topologies. Working with spectral spaces equipped with both these topologies and the corresponding notion of frames, one gets a natural bitopological description of coherent morphisms between coherent frames, and thus of proper maps between spectral spaces, see Picado (1994), Jung and Moshier (2006), and Bezhanishvili et al. (2010). As an alternative to restricting the morphisms on spectral spaces and coherent frames to fit those of bounded distributive lattices, we can weaken the notion of morphism on DL to correspond to lattice homomorphisms A → Idl(B), which in turn may be seen as certain “join-approximable” relations from A to B (see Exercise 6.3.7). Exercises for Section 6.3 Exercise 6.3.1 Give an example of an infinite frame in which each element is a finite meet of meet-irreducible elements and one in which it is not the case. Exercise 6.3.2 This exercise contains material beyond the scope of this book. It is given mainly as an indication of a means of getting one’s hands on non-spatial frames. (a)

(b)

Show that non-atomic complete Boolean algebras are examples of non-spatial frames. To this end you may proceed as follows. (1) Every complete Boolean algebra is a frame, that is, it satisfies (JID). (2) A complete Boolean algebra is a spatial frame if, and only if, it is atomic, that is, for all a ∈ B, if a  ⊥, then there exists an atom j ∈ B such that j ≤ a. Show that there exist non-atomic complete Boolean algebras. Hint. We outline two rather distinct ways of proving this. (1) Show that there exist non-atomic Boolean algebras (for example, the Lindenbaum–Tarski algebra of Classical Propositional Logic on a countable set of primitive propositional variables, see Section 4.1, has no atoms at all), and show that the MacNeille completion (MacNeille, 1937) of a Boolean algebra is atomic if, and only if, the original Boolean algebra is atomic; or (2) Show that the regular open subsets of a

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

210

6 Omega-Point Duality

Hausdorff space form a complete Boolean algebra, and that this algebra has no atoms if the original space has no isolated points. Recall that an open subset in a topological space is called regular if it is equal to the interior of its closure. Exercise 6.3.3 (a) Show that all Hausdorff spaces are sober. (b) Let X be an infinite set equipped with the topology in which a subset is closed if, and only if, it is finite or equal to X. Prove that this space is T1 and not sober. Exercise 6.3.4 Let X be a topological space. Prove that the closure of a singleton set is join irreducible in the lattice of closed subsets of X. Exercise 6.3.5 Show that if X is a locally compact T0 space, then X is well filtered if, and only if, it is sober. Use this in combination with Exercise 6.1.3 to show that a topological space is a spectral space if, and only if, the following three properties hold: (a) (b) (c)

The collection of compact-open subsets of X forms a bounded sublattice of P (X ) and a base for the topology of X; X is sober; and X is T0 .

Exercise 6.3.6 Let L be a frame. (a) (b) (c)

(d)

Show that for x, y ∈ Pt(L) we have x ≤ y in the specialization order if, and only if, f x ≤ f y if, and only if, F (x) ≤ F (y). Show that binary joins and meets need not exist in the specialization order in a sober space. Show that the poset of completely prime filters of a frame is closed under directed union. Conclude that any sober space is a dcpo in its specialization order. Show that each open of Pt(L) is Scott open with respect to the specialization order of Pt(L).

Note. Item (c) of this exercise has a counterpart for distributive lattices and spectral spaces (see Exercise 3.1.12). Also, in Proposition 7.6 we give a direct, elementary proof of the results in this exercise that does not use Ω-Pt duality. Exercise 6.3.7 Let L and M be distributive lattices. A relation R ⊆ L × M is called join approximable (see, for example, Abramsky and Jung (1994, Definition 7.2.24)) provided that, for any a, a  ∈ L and b, b ∈ M, the following four properties hold: • if a  ≥ aRb ≥ b, then a  Rb; • if aRb and aRb, then aRb ∨ b;

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

211

• if aRb and a  Rb, then a ∧ a  Rb; and • if a ∨ a  Rb, then there exist c, c  ∈ K such that aRc, a  Rc , and b ≤ c ∨ c . Denote by X and Y the spectral spaces dual to L and M, respectively. (a)

(b) (c)

Show that there is a one-to-one correspondence between continuous functions f : Y → X and DL homomorphisms h : L → Idl(M). Hint. Use the fact that Ω(Y ) is isomorphic to Idl(M). Show that there is a one-to-one correspondence between DL homomorphisms h : L → Idl(M) and join-approximable relations R ⊆ L × M. Conclude that the category whose objects are distributive lattices and whose morphisms are join-approximable relations is dually equivalent to the category Specc of spectral spaces with continuous functions between them.

Note. A related result on algebraic domains will be stated in Exercise 7.1.12. 6.4 Duality for Spaces of Relations and Functions As an application of the duality between distributive lattices and spectral spaces, in this section, we will develop a general duality theory for spaces of relations and functions, which we will apply to domains in Chapter 7. We begin in the setting of general spectral spaces, and we will start with a class of constructors, on the lattice side, which are central in logic applications as already considered in Chapter 4, namely that of freely adding one layer of a modal box, and of an implicationtype operator. We will see in particular that the dual of a layer of implication-type operator gives a “binary relation space” (Theorem 6.28). In order to obtain from this binary relation space a function space, we identify (Definition 6.32) a property that we call “preserving joins at primes,” which we will encounter again in Chapters 7 and 8. We prove (Corollary 6.34) that, for certain lattices, this property gives a lattice-theoretic description of the function space construction, which will allow us to prove in Chapter 7 that bifinite domains are closed under that construction. Freely Adding a Layer of Unary Modal Operator An important idea from categorical algebra is that the addition of algebraic structure to an object A in a category C can sometimes be understood as a homomorphism T A → A, where T is an endofunctor on C. In particular, Exercise 5.4.3 spells out how a positive modal -algebra ( A, ) can alternatively be specified by giving a distributive lattice A and a lattice homomorphism F ( A) → A. Here, we can think of F as “freely adding one layer of unary  operator” to a given distributive lattice. Thus, an important step in analyzing positive modal -algebras is to analyze this “signature” functor F itself. In this subsection, we will identify the topological

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

212

6 Omega-Point Duality

dual of this functor, and, in the next subsection, we carry out a similar program for the functor F→ , which “freely adds a layer of implication-type operator” to a distributive lattice. Before treating general implication-type operators, we first consider the case of a unary modal operator, which, as we will see, corresponds to the filter space construction. We also show how this construction specializes to the upper Vietoris space, which is also known as the Smyth powerdomain in the context of the spectral domains that we discuss in Chapter 7. We recall the construction of the free distributive lattice over a meet-semilattice (also see Example 5.42 and Exercise 5.4.3 for more details). Let M be a meetsemilattice. The distributive lattice F (M) is characterized uniquely by the universal property that any finite-meet-preserving function M → L, with L a distributive lattice, lifts uniquely to a DL homomorphism F (M) → L. Recall from Example 5.42 that we may construct F (M) as the quotient of the free distributive lattice FDL (M) over the set of “formal boxes” {a | a ∈ M } under the congruence ϑ M generated by the set of pairs of the form  

aG ≈ bG, where aG :=  G and bG := {a : a ∈ G}, (6.1) and where G ranges over the finite subsets of M. We follow here the same convention as in Section 4.2, that the pairs involved in generating a congruence are denoted by a ≈ b instead of (a, b). We now give a method for calculating the Stone–Priestley dual space of this lattice F (M), which we will subsequently generalize to the setting of implication-type operators. Denote by X the Priestley dual space of F (M) and by Y the Priestley dual space of FDL (M). Recall from Proposition 4.8 that we may regard the space Y as the ordered generalized Cantor space over M, as given in Definition 4.7; that is, the underlying set of Y is 2M , the partial order Y on Y is the pointwise order induced by the order on 2 in which 1 ≤ 0, that is, for y, y  ∈ Y , def

y Y y  ⇐⇒ for all m ∈ M, if y  (m) = 1, then y(m) = 1 , and the Priestley topology on Y is generated by the subbase of sets of the form m → i = {y ∈ 2M | y(m) = i}, where m ∈ M and i ∈ {0, 1}. Since F (M) is the quotient of FDL (M) by ϑ M , by Theorem 4.17, the Priestley space X is order-homeomorphic to the closed subspace of Y consisting of the points of Y that “respect” all the pairs generating ϑ M ; using the notation of Theorem 4.17 and (6.1), X is order-homeomorphic to {aG ≈ bG | G ⊆ M finite }. Now note that, for any finite subset G ⊆ M, a point y ∈ 2M satisfies the equation aG ≈ bG if, and only if,   y(a). y  G = a ∈G

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

213

In other words, the points y ∈ Y that satisfy all equations of the form (6.1) are exactly the points y ∈ 2M that preserve finite meets when viewed as functions M → 2, by precomposing with the bijection a → a. Finally, note that the meet-preserving functions M → 2 correspond exactly to the filters of M, by considering the inverse image of 1 and using the same definition of “filter,” now for a meet-semilattice, as in Definition 3.2. Let us write ϕ for the induced bijection X → Filt(M), which can be defined concretely, for x ∈ X, by ϕ(x) := {m ∈ M | x(m) = 1}. Under this bijection ϕ, the partial order X on the space X, inherited from the order Y on Y , corresponds to the reverse inclusion on Filt(M), that is, for any x, x  ∈ X, x X x  if, and only if, ϕ(x) ⊇ ϕ(x  ). Since the set M := {[m]ϑ M | m ∈ M } generates F (M) as a lattice, the lattice of compact-opens of the spectral topology τ ↓ a, as a ranges over the elements of M. on X is also generated by the sets of the form  a is the set Note that, for any a ∈ M, the direct image under the bijection ϕ of the set   a := {F ∈ Filt(M) | a ∈ F}. Note also that the collection of sets  a, as a ranges over the elements of M, is closed  and  under finite intersections, since Filt(M) =  a ∩ b = a ∧ b for any a, b ∈ M. Denote by ρ the topology on Filt(M) generated by the base { a | a ∈ M }. It then follows from the above calculations that the function ϕ, viewed now as a map between topological spaces (X, τ ↓ ) → (Filt(M), ρ), is a homeomorphism. Therefore, ρ is a spectral topology, and F (M) is isomorphic to the lattice of compact-opens of ρ, via the unique homomorphism extending the function which sends a to a, ˜ for a ∈ M. In particular, each set of the form  a is compact-open in ρ. Note also that any compact-open set in (Filt(M), ρ) is a finite union of sets from this base, since any open set in ρ is a union of sets from the base. We summarize our findings in the following proposition. Proposition 6.20 Let (M, ∧, ) be a meet-semilattice. The spectral space dual to F (M) is homeomorphic to the set of filters Filt(M) of M, equipped with the topology generated by the base { a | a ∈ M }. Moreover, the compact-open sets of Filt(M) are exactly the finite unions of sets in this base. To finish this first subsection, suppose that M, in addition to being a meetsemilattice, is itself also a distributive lattice. One may then directly describe the space of filters Filt(M) in terms of the spectral space dual to M, using the following general construction. Definition 6.21 Let X be a topological space. The upper Vietoris space on X is the topological space V ↑ (X ) whose points are the compact-saturated subsets of X, and whose topology is generated by the base consisting of the sets

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

214

6 Omega-Point Duality

U := {K ∈ V ↑ (X ) | K ⊆ U}, for U ∈ τ. Note that, for K, K  ∈ V ↑ , we have K ≤ K  in the specialization order if, and only if, K  ⊆ K. In the context of general non-Hausdorff topological spaces, the name “Vietoris hyperspace” has been associated to various spaces whose points are the subsets of a given topological space, with a certain topology induced from that space. Often, the points of such a hyperspace are the closed sets, and indeed, in our definition here, if X is stably compact, then the points of V ↑ (X ) are in fact the closed subsets of the space X ∂ , the co-compact dual of X. Proposition 6.22 Let M be a distributive lattice and let X be the spectral space of M. The topological space Filt(M) is homeomorphic to the upper Vietoris space V ↑ (X ). Proof Let f : V ↑ (X ) → Filt(M) be the function that sends a compact-saturated set K of X to the filter {a ∈ M | K ⊆  a }. We will show that this function f is continuous and has a continuous two-sided inverse, g : Filt(M) → V ↑ (X ), which  sends a filter F of M to the set { a | a ∈ F}, which is indeed compact and saturated in X. Now, to see that g is a two-sided inverse to f , let F be a filter of M and b ∈ M. Then, using well-filteredness of the spectral space X, g(F) ⊆  b if, and only if, there exists a ∈ F such that  a ⊆  b, which is clearly equivalent to saying that b ∈ F. Thus, f g(F) = F. Conversely, for any compact-saturated set K of X, we have  g f (K ) = { a | K ⊆ a } = K, as can be seen from the fact that X ∂ is a spectral space in which the sets  a c form a base for the open sets. To see that f is continuous, let a ∈ M be arbitrary. For any K ∈ V ↑ (X ), we have a ) =  a, which is f (K ) ∈  a if, and only if, K ⊆  a if, and only if, K ∈  a, so f −1 ( ↑ open in V (X ). Finally, to see that g is continuous, let U be a basic open set of V ↑ (X ), where U is an arbitrary open in X. Let F be a filter of M and suppose that g(F) ∈ U. We show that there exists an open set around F which is mapped entirely inside U by  g. By definition of g and of U, we have g(F) = { a | a ∈ F} ⊆ U, and since X

 is a spectral space, we may write U as the union { b : b ∈ M,  b ⊆ U}. Using the well-filteredness of X, pick some a ∈ F such that  a ⊆ U. Then, F ∈  a, and for any     a, we have g(F ) ⊆ b ⊆ U, so g(F ) ∈ U.  F ∈ Note in particular that it follows from the above results that spectral spaces are closed under the upper Vietoris construction. Indeed, for a spectral space X, writing M := K O(X ), the space V ↑ (X ) is homeomorphic to Filt(M) by Proposition 6.22, which is the dual space of F (M), and is therefore a spectral space.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

215

Corollary 6.23 Spectral spaces are closed under the upper Vietoris construction.

Freely Adding a Layer of Implication-Type Operator The above calculation of the dual space of F (M) suggests the following general methodology, which we will follow here to construct function spaces via their dual lattices. Suppose that a lattice K is described by a set of generators V and a set of equalities R between distributive lattice terms in the generators, and that we want to compute its Priestley dual space (X, ≤, τ). One starts from a free distributive lattice over V , whose dual space is 2V . Then, every equation a ≈ b ∈ R between generators yields a quotient of FDL (V ) and thus a closed subspace of 2V . Intersecting all these subspaces, as a ≈ b ranges over the equations in R, yields the dual space X of the distributive lattice in question. This is an application of the quotient-lattice– subspace duality of Section 4.2, where algebraic equations yield subspaces. An interesting point of this method is that one may prove properties of this dual space by examining the corresponding properties of the lattice; for example, the space (X, τ ↓ ) will automatically be a spectral space, by virtue of being the dual space of a distributive lattice. This method will be used in our study of domain theory in Section 7.4, and has also been used in the study of free algebras in varieties of lattice ordered algebras, see for example Ghilardi (1992, 2010), Bezhanishvili and Gehrke (2011), or Coumans and van Gool (2012). We now apply the same methodology to implication-type operators. Let L and M be distributive lattices and X and Y their respective dual spaces. We will think of the free distributive lattice over the set L × M as generated by “formal implications” a→b for a ∈ L and b ∈ M. For this reason we will here denote elements of L × M, when we consider them as generators of the free distributive lattice, by a → b rather than the usual (a, b). We aim to calculate (see Theorem 6.28) the dual space of the quotient F→ (L, M) of FDL (L × M) given by the congruence generated by the following two sets of equations, which we also refer to as (equational) “schemes” in what follows:   a→ G≈ {a → b | b ∈ G}, for each a ∈ L and finite G ⊆ M and (6.2) 

F→b≈



{a → b | a ∈ F}, for each finite F ⊆ L and b ∈ M.

(6.3)

Remark 6.24 Notice that the schemes (6.2) and (6.3) together precisely say that →, viewed as an operation from L × M to the algebra we are building, is an implication-type operator, in the sense of Definition 4.58 in Section 4.5. Note

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

216

6 Omega-Point Duality

also that the implication → that we build here will not in general be a Heyting implication, that is, it will not be adjoint to a meet operation ∧. Also, as in the unary case, F→ corresponds to freely adding a layer of implicationtype operator in the sense that the coproduct L + F→ (L, L) is the sublattice of depth ≤ 1 terms in the free distributive lattice with an implication-type operator over L (see Exercise 5.4.3). Again using Proposition 4.8, as we did in the case of a unary operator above, we identify the points of the dual space of the free distributive lattice on L × M with subsets of L × M. Concretely, any prime filter F of FDL (L × M) gives a subset SF = F ∩ (L × M) of L × M, and conversely, if S is any subset of L × M, then the filter FS of FDL (L × M) generated by S is prime. Indeed, by Exercise 3.1.13, we have    T ≤ u for some finite T ⊆ S , FS = u ∈ FDL (L × M) |  and this filter is prime: for any T ⊆ S finite, T is join prime in FDL (L × M). That n ui , is, in disjunctive normal form, any u ∈ FDL (L × M) is of the form u = i=1 where each ui is the meet of a finite set of generators Ui ⊆ L × M and we have n    T ≤ ⇐⇒ T ⊇ Ui for some i with 1 ≤ i ≤ n. Ui i=1

Under this correspondence, if S ⊆ L × M is any subset and t, u ∈ FDL (L × M) is a pair of lattice terms in the set of generators {a → b : a ∈ L, b ∈ M }, then we say S satisfies the equation t ≈ u when the prime filter FS is in the subspace t ≈ u defined by the equation t ≈ u, in the sense of the quotient-lattice–subspace duality of Theorem 4.27. In other words, saying that S satisfies t ≈ u simply means that t ∈ FS if, and only if, u ∈ FS . The dual space of the congruence generated by a relation R on FDL (L × M) may then be identified with the collection of subsets S ⊆ L × M that satisfy all the equations in R. In particular, in the following lemma we calculate the subspaces of 2 L × M corresponding to the schemes (6.2) and (6.3). For any S ⊆ L × M, a ∈ L, and b ∈ M, we define Sa := {b ∈ M | a → b ∈ S} and S b := {a ∈ L | a → b ∈ S}. Lemma 6.25 A subset S ⊆ L × M satisfies (6.2) if, and only if, for each a ∈ L, Sa is a filter of M and S ⊆ L × M satisfies schema (6.3) if, and only if, for each b ∈ M, S b is an ideal of L. Proof We prove just the first statement as the second follows by symmetry and   order duality. Since a → G is a generator, a → G ∈ FS if, and only if,   a → G ∈ S, while {a → b | b ∈ G} ∈ FS if, and only if, {a → b | b ∈ G} ⊆ S. That is, S satisfies all equalities in (6.2) if, and only if, for each a ∈ L, the set Sa  has the property that, for any finite subset G of S, G is in Sa precisely when G is  a subset of Sa . That is, if, and only if, Sa is a filter of M (see Exercise 3.1.1).

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

217

We will establish, in Theorem 6.28, a first connection between the implicationtype operator and the function space construction. To do so, we need the following general definition of a topology on a function space. Definition 6.26 Given two topological spaces Z1 and Z2 , we denote the set of continuous functions from Z1 to Z2 by [Z1, Z2 ]. The compact-open topology 2 on [Z1, Z2 ] is defined to be the topology generated by the subbase consisting of the sets K → U := { f ∈ [Z1, Z2 ] | f [K] ⊆ U} for K ∈ K S(Z1 ) and U ∈ Ω(Z2 ), where we recall that K S(Z1 ) denotes the collection of compact-saturated subsets of Z1 and Ω(Z2 ) denotes the collection of open subsets of Z2 . See Exercise 6.4.3 for more detail and some basic properties used in the proofs below. Lemma 6.27 Let L be a distributive lattice with dual spectral space X and let M be a meet-semilattice. Then, the compact-open topology on [X, Filt(M)] is generated by the subbase consisting of the sets  a→ b for a ∈ L and b ∈ M. Proof By Proposition 6.20, the space Filt(M) with the topology generated by the sets  b for b ∈ M is a spectral space, as it is the dual of F (M). Thus it follows, by Exercise 6.4.3, that the sets  a → V for a ∈ L and V compact-open in Filt(M) form a subbase for the topology on [X, Filt(M)]. The compact-opens of Filt(M) are finite unions of sets of the form  b for b ∈ M. Thus, we must show that such sets are in the topology generated by the  a→ b with a ∈ L and b ∈ M. To this end, let  a ∈ L and G ⊆ M be finite. Define C := {c ∈ L G | a ≤ g ∈G cg }. We will prove that ⎫ ⎧  ⎪ ⎪ ⎪ ⎪ ⎬. ⎨     ( c → g ) | c ∈ C g= a→ g ⎪ ⎪ ⎪ ⎪ g ∈G ⎭ ⎩ g ∈G For the right-to-left inclusion, let c ∈ C be arbitrary and suppose that f ∈   g ). For any x ∈  a, since a ≤ g ∈G cg , there is g ∈ G with x ∈ cg , g ∈G ( cg → 

g it follows that f (x) ∈  g . Thus, f ∈  a → g ∈G  g. and since f ∈ cg → 

For the converse, let f ∈  a → g ∈G  g . For every x ∈  a, pick gx ∈ G with f (x) ∈ gx . Since f is continuous, f −1 (gx ) is open. Since X is a spectral space, pick

a x ⊆ f −1 (gx ). Then,  a ⊆ x ∈ a x , and thus, by compactness of a x ∈ L with x ∈  a

 a, there is a finite F ⊆  a with  a ⊆ x ∈F  ax . 2

Note that the adjective compact-open in “compact-open topology,” which is common in the literature, refers to the fact that the subbase is given by sets of functions determined by a compact and a (different) open set. This contrasts with the use of “compact-open set” throughout this book, meaning a single set that is compact and open, also common in the literature.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

218

6 Omega-Point Duality

 Now, for each g ∈ G, define cg := {a x | x ∈ F and gx = g}. Then, since each gx ∈ G, we obtain   

 {a x | x ∈ F and gx = g} = ax = cg . a≤ x ∈F

g ∈G

g ∈G

That is, c = (cg )g ∈G ∈ C. Also, for each x ∈  a, we have  a x ⊆ f −1 (gx ), or ax →  gx . Therefore, we obtain equivalently, f [ a x ] ⊆ gx , so f ∈ 

{ ax →  ( ax →  gx ) = g | x ∈ F and gx = g} f ∈ x ∈F

g ∈G

=





{ a x | x ∈ F and gx = g} →  g = ( cg →  g ),

g ∈G

g ∈G

where the second equality holds since →, with the second argument fixed, sends finite unions in the first argument to finite intersections (see Exercise 6.4.3(b)(1)). Note that this is part of saying that → is an operation of implication type.  Using our results above, we will now give a concrete description of the dual space of F→ (L, M) for arbitrary distributive lattices. In the following theorem, we will show that, viewed on the dual spectral spaces X and Y , the construction F→ corresponds to a space of upward Priestley compatible relations from X to Y , also see Exercise 6.4.1. Theorem 6.28 Let L and M be distributive lattices and let X and Y be their respective dual spectral spaces. The spectral space dual to the distributive lattice F→ (L, M) := FDL (L × M)/ϑ, where ϑ is the congruence generated by the schemes (6.2) and (6.3) together, is homeomorphic to the space [X, V ↑ (Y )], of continuous functions from X to the upper Vietoris space of Y , equipped with the compact-open topology. Proof By Proposition 6.22, the space V ↑ (Y ) is homeomorphic to Filt(M). It is therefore equivalent to prove that the dual space of F→ (L, M) is homeomorphic to the space [X, Filt(M)] with the compact-open topology. Also, by Lemma 6.25 and the considerations preceding it, the dual space of F→ (L, M) is the subspace Z of 2 L×M consisting of those subsets S of L × M for which Sa is a filter of M for each a ∈ L and S b is an ideal of L for each b ∈ M. To prove the theorem, we will define a mutually inverse pair of continuous maps ϕ : Z  [X, Filt(M)] : ψ. For S ∈ Z, define a function ϕ(S) by  ϕ(S) : X −→ Filt(M), x → {Sa | a ∈ Fx }. We need to show that ϕ(S) is a well-defined, continuous function from X to Filt(M). To see that the function ϕ(S) is well defined – that is, that ϕ(S)(x) is a filter for every x ∈ X – it suffices to show that the collection of filters {Sa | a ∈ Fx } is directed, by Exercise 3.1.12. To this end, note that the assignment a → Sa is order

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

219

reversing; indeed, if b ∈ Sa , then a  ∈ S b , so a ∈ S b since S b is a down-set, so that b ∈ Sa . Thus, since Fx is down-directed, the collection {Sa | a ∈ Fx } is up-directed, as required. We now show that ϕ(S) is continuous. Let b ∈ M. For every x ∈ X, we have b) if, and only if, there exists a ∈ Fx such that a → b ∈ S; in a formula: x ∈ ϕ(S) −1 (   b) = c. (6.4) ϕ(S) −1 ( c ∈Sb

b) is open for every b ∈ M, so ϕ(S) is a continuous function In particular, ϕ(S) −1 ( from X to Filt(M). For the inverse function ψ, let f : X → Filt(M) be continuous. Define ψ( f ) := {a → b ∈ L × M |  a ⊆ f −1 ( b)}. We first show that ψ( f ) ∈ Z. Note that, for any b ∈ M, a ⊆ f −1 ( b)}, ψ( f ) b = {a ∈ L |  which is an ideal because  ( ) preserves finite joins. Similarly, for any a ∈ L, a] ⊆  b}, ψ( f )a = {b ∈ M | f [ and since  ( ) preserves finite meets, this is a filter. So, ψ( f ) ∈ Z. To show that ϕ and ψ are mutually inverse, first note that, for any b ∈ M, f ∈ [X, Filt(M)] and x ∈ X, we have b ∈ ϕ(ψ( f ))(x) ⇐⇒ ∃a ∈ L

and

b) ⇐⇒ b ∈ f (x), x∈ a ⊆ f −1 (

where we use in the second equivalence that f is continuous. Thus, ϕ(ψ( f )) = f . Also, for any S ∈ Z, a ∈ L, and b ∈ M, we have   b) = c, a → b ∈ ψ(ϕ(S)) ⇐⇒  a ⊆ ϕ(S) −1 ( c ∈S b

recalling (6.4). By compactness of X and the fact that  ( ) is an order embedding, the latter inclusion is equivalent to: There exists a finite subset C of S b such that  a ≤ C. Since S b is an ideal, this is in turn equivalent to a ∈ S b , that is, a → b ∈ S, as required. It remains to prove that the bijections ϕ and ψ between Z and [X, Filt(M)] are homeomorphisms. Observe that a basic open of Z, which is of the form a → b = {S ∈ Z | a → b ∈ S}, is sent by ϕ to the set  a→ b = { f ∈ [X, Filt(M)] |  a ⊆ f −1 ( b)}. By Lemma 6.27, these sets form a base for [X, Filt(M)] with the compact-open topology, and thus both ϕ and ψ are continuous. 

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

220

6 Omega-Point Duality

We may recover the upper Vietoris space construction itself as a special case of the “relation space” [X, V ↑ (Y )] used in Theorem 6.28. Indeed, notice that V ↑ (X )  [X ∂, S]  [X ∂, V ↑ (1)], where S is the Sierpinski space and 1 is the one-element space (see Exercise 6.4.2). Remark 6.29 Let X and Y be spectral spaces with dual lattices L and M, respectively. Then, we have shown that [X, V ↑ (Y )] is again a spectral space. Even though V ↑ (Y ) is dual to the lattice F (M), obtained by freely adding a layer of unary dual operator to M, notice that the compact-opens are not given by the elements of L and F (M), but rather by the elements of L and M, see Example 6.35. Remark 6.30 Notice that in Theorem 6.28 we are considering only operators of implication type but, with some order flips, this result may be transposed to other types of operators. For example, consider the construction that takes distributive lattices L and M and produces a distributive lattice F• (L, M) by freely adding a binary operation • that preserves finite meets in both coordinates; that is, keeping the   analogue for • of scheme (6.2) and replacing (6.3) by ( F) • b ≈ {a • b | a ∈ F} for each finite F ⊆ L and b ∈ M. The dual space of this construction is then [X ∂, V ↑ (Y )], where X ∂ denotes X equipped with the co-compact dual of the topology of X.

Preserving Joins at Primes Let X and Y be spectral spaces with dual distributive lattices L and M, respectively. We are interested in the space of continuous functions [X, Y ], which can be regarded as a (generally non-spectral) subspace of the space of compatible relations [X, V ↑ (Y )], as we explain now (see Exercise 6.4.4 for more details). Denote by η : Y → V ↑ (Y ) the embedding of Y in V ↑ (Y ) given by y → ↑y. Then, we have an injective function [X, Y ] → [X, V ↑ (Y )] by sending f ∈ [X, Y ] to η ◦ f , and a base for the topology on [X, Y ], now viewed as a subspace of [X, V ↑ (Y )], is given by the sets  a→ b = ( a→ b) ∩ [X, Y ], for a ∈ L and b ∈ M. Even though Y is a Priestley-closed subspace of V ↑ (Y ), it is not in general the case that [X, Y ] is a Priestley-closed subspace of [X, V ↑ (Y )], reflecting the fact that [X, Y ] is not in general a spectral space; we give an explicit example of this occurrence in Example 6.35. One would need to move to frames, sober spaces, and geometric theories to describe [X, Y ] as the dual of a quotient. However, we have a finitary approximation of the subspace [X, Y ], that we will give in Theorem 6.31. Here, recall from Section 4.2 that if Z is a closed subspace of a Priestley space P, then the corresponding congruence ϑ on the dual lattice A of P is given by a∩Z = b ∩ Z }, ϑ = {(a, b) ∈ A2 | 

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

221

so that, for any a, b ∈ A, we have [a]ϑ ≤ [b]ϑ if, and only if, for every z ∈ Z, if z ∈  a, then z ∈  b.

(6.5)

We apply this now in the case where P is the Priestley space associated to the spectral space [X, V ↑ (Y )], whose topology, we recall, is the patch topology of the compact-open topology and its co-compact dual. Theorem 6.31 Let L and M be distributive lattices, X and Y their respective dual spaces, and Z a subset of [X, V ↑ (Y )] which is closed in the patch topology. Denote by ϑ the congruence on F→ (L, M) corresponding to Z. The following are equivalent: (i) (ii)

Z is a subset of [X, Y ]; and for every x ∈ X, a ∈ Fx , and finite subset G ⊆ M, there is c ∈ Fx such that      a→ G ≤ {c → b | b ∈ G} . (6.6) ϑ

ϑ

Proof In light of Proposition 6.22, we will work with Filt(M) instead of V ↑ (Y ), and we may consider Z as a subset of [X, Filt(M)]. As explained above, under the identification V ↑ (Y )  Filt(M), the subspace Y of V ↑ (Y ) corresponds to the subspace PrFilt(M) of prime filters of M, so [X, Y ] corresponds to [X, PrFilt(M)]. First, suppose that (i) holds, that is, Z ⊆ [X, PrFilt(M)]. Let x ∈ X, a ∈ Fx , and G ⊆ M finite. We need to show that there exists c ∈ Fx such that (6.6) holds. We  will first show that, for any f ∈ Z with f ∈ a → G, there exist c f ∈ Fx and b f ∈ G such that f ∈ c f → b f . To see this, let f ∈ Z be arbitrary and suppose   that f ∈ a → ( G). Then, by definition, f [ a] ⊆  G and thus, as x ∈  a, we have  G ∈ f (x). Now, since f ∈ Z ⊆ [X, PrFilt(M)], the filter f (x) is prime, so we b. Since f is continuous, may pick b f ∈ G with b f ∈ f (x), or equivalently, f (x) ∈  pick c f ∈ L with x ∈ cf and f [cf ] ⊆ bf , or equivalently, f ∈ cf → bf = c f → bf .  Now, the sets c f → b f , for f ranging over the patch-closed set Z ∩ a → ( G), are a cover of this set. By compactness of the patch topology, pick a finite subcover, n c fi . Then, c ∈ Fx since each c fi is indexed by f 1, . . . , f n , say, and define c := i=1  in Fx . Let us show that for this c, (6.6) holds. Writing d := {c → b | b ∈ G},   → G, then we want to show that [a → G]ϑ ≤ [d]ϑ . For any f ∈ Z, if f ∈ a  f ∈ c fi → b fi for some 1 ≤ i ≤ n. It now follows that f ∈ d, since c fi → b fi ≤ c → b fi ≤ d, where we use first that → is order reversing in the first coordinate, and then that   → G, then f ∈ d. b fi ∈ G. Thus, we have shown that for any f ∈ Z, if f ∈ a Using (6.5), we thus conclude (6.6). Conversely, suppose (ii) holds. Let f ∈ [X, Filt(M)] be in Z and x ∈ X. If  G ⊆ M is finite and G ∈ f (x), then, as f is continuous, there is a ∈ L with

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

222

6 Omega-Point Duality

   f [ a] ⊆  G, or equivalently, f ∈  a →  G = a→ ( G). Now, using (6.6), it → b. Thus, x ∈  c and f [ c] ⊆  b follows that there are c ∈ Fx and b ∈ G with f ∈ c and thus b ∈ f (x). That is, we have shown that f (x) is a prime filter and thus that f ∈ [X, PrFilt(M)]  [X, Y ], as required.  Definition 6.32 Let K, L, and M be distributive lattices and let → : L × M → K be an implication-type operator. We say that → preserves joins at primes if, for every prime filter F of L, a ∈ F, and finite subset G of M, there exists c ∈ F such that   a→ G≤ {c → b | b ∈ G}. For a congruence ϑ on F→ (L, M), we also say that ϑ makes → preserve joins at primes or → preserves joins at primes modulo ϑ if the equivalent properties in Theorem 6.31 hold. Remark 6.33 To explain the above terminology, note that the property of preserving joins at primes modulo ϑ is equivalent to the property that, for each x ∈ X, the following operation preserves finite joins: x → (−) : M → Idl(F→ (L, M)/ϑ) b →  → b]ϑ | a ∈ xIdl . In a lattice with enough join primes, there is actually a largest congruence that makes → preserve joins at primes, see Corollary 6.34. This congruence is used crucially in Theorem 7.69 in Chapter 7. The property of preserving joins at primes is closely related to being determined by finite quotients. See Theorem 8.45 in Chapter 8 and Gehrke (2016, Theorem 3.18) for another occurrence of this phenomenon in the setting of topological algebras on Boolean spaces. For a study of this notion via canonical extensions, see Fussner and Palmigiano (2019). Recall that an element p in a lattice L is said to be join prime provided that, for any  finite F ⊆ L, p ≤ F implies that there exists a ∈ F with p ≤ a. Further, we say that L has enough join primes provided every element of L is the join of a finite set of join-prime elements. Also, recall from Chapter 1 that we denote by J (L) the poset of join-prime elements of L, with order inherited from L. The following special case of Theorem 6.31, where one of the two lattices is assumed to have enough join primes, is central to the treatment of the function space operator in Domain Theory in Logical Form (Abramsky, 1991), as we will also see in Section 7.4. Corollary 6.34 Let L and M be distributive lattices and suppose L has enough join primes. Further, let X and Y be the dual spectral spaces of L and M, respectively. The space [X, Y ] of continuous functions from X to Y equipped with the compactopen topology, is dual to the congruence ϑjpp on F→ (L, M) that is generated by the following set of equations:

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

p→



G≈



223

{p → b | b ∈ G},

(6.7)

where p ranges over the join-prime elements of L and G ranges over the finite subsets of M. Proof We first note that it suffices to prove that a congruence ϑ of F→ (L, M) makes → preserve joins at primes if, and only if, ϑjpp is contained in ϑ. Indeed, by Theorem 6.31, this claim implies that ϑjpp is the minimum congruence whose dual closed set is contained in [X, Y ], and thus the dual Zjpp of F→ (L, M)/ϑjpp is the maximum patch-closed subspace of [X, Y ]. If Zjpp were a proper subset of [X, Y ], then there would be f ∈ [X, Y ] – Zjpp , and Zjpp ∪ { f } would still be patch closed, since singletons are closed in the patch topology. Thus, Zjpp must be the full subspace [X, Y ]. To prove that ϑjpp is indeed the minimum congruence that makes → preserve joins at primes, suppose first that ϑ is a congruence which makes → preserve joins at primes and let p ∈ J (L) and G ⊆ M finite. Instantiating the condition (6.6) for Fx = ↑p and a = p, pick an element c ∈ ↑p so that      p→ G ≤ {c → b | b ∈ G} . ϑ

ϑ

Now, p ≤ c implies that c → b ≤ p → b in F→ (L, M), so the right-hand side of the  above inequality is at most [ {p → b | b ∈ G}]ϑ . We conclude that      p→ G ≤ {p → b | b ∈ G} . ϑ

ϑ

The other inequality holds because → is order preserving in its second coordinate. Thus, ϑjpp is contained in ϑ. Conversely, suppose ϑ is a congruence that contains ϑjpp and let x ∈ X, a ∈ Fx , and G ⊆ M be finite. Since L has enough join primes, there is a finite set F ⊆ J (L)  so that a = F. Also, since a ∈ Fx and Fx is a prime filter, there is p ∈ F with p ∈ Fx . Further, as the function → from L × M to F→ (L × M) is an implication-type   operator, we have a → G ≤ p → G in F→ (L × M), since p ≤ a. It follows that         a→ G ≤ p→ G = {p → b | b ∈ G} , ϑ

ϑ

ϑ

using the assumption that ϑ contains ϑJPP . Now, in the notation of Definition 6.32, taking c := p shows that ϑ makes → preserve joins at primes.  We finish this subsection by giving the promised example that [X, Y ] itself is not always a Priestley-closed subspace of [X, V ↑ (Y )]. Example 6.35 Let L be ⊥ ⊕ (Nop × 2) and X its dual spectral space, both as depicted in Figure 6.3. Here, x i = ↑ai is a principal prime filter for each i ∈ N, while x = L –{⊥} and y = {bi | i ∈ N}. Further, we let Y be the finite spectral space

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

224

6 Omega-Point Duality b0 = 

x

a0 b1 a1 y

b2 x2

a2

x1 ⊥

x0

Figure 6.3 The lattice L from Example 6.35 and the specialization order of its dual spectral space X. y0

y1

y2

Figure 6.4 The space Y from Example 6.35.

depicted in Figure 6.4. Note that [X, Y ] is not compact since we have the following infinite cover, that clearly cannot have a finite subcover:   [X, Y ] = (X → ↑y1 ) ∪ (X → ↑y2 ) ∪ bi → ↑y2 ) ( a0 → ↑y1 ) ∩ ( i ∈N

∪



 bi → ↑y1 ) . ( a0 → ↑y2 ) ∩ (

i ∈N

This equality is saying that a continuous function f from X to Y either misses y1 or it misses y2 and otherwise x 0 gets sent to y1 or y2 and y gets sent to the other. In these last cases, only finitely many of the x i take the same value as x 0 and this means that there exists n ∈ N such that the direct image of  an under f is {y0 }. Compare also to the proof of Lemma 6.27.

Exercises for Section 6.4 Exercise 6.4.1 Let M be a distributive lattice with dual Priestley space (X, ρ, ≤) and let (V ↑ (X ), τ) be the upper Vietoris space of the spectral space (X, ρ↑ ). This

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

6.4 Duality for Spaces of Relations and Functions

225

exercise makes a link between the calculation of the dual space of the lattice F (M) and the duality for unary operators  of Section 4.3. (a) (b)

(c)

Prove that τ ∂ , the co-compact dual topology of τ, is generated by the base consisting of finite unions of sets of the form (a) c , for a ∈ M. Conclude that the Priestley dual space of F (M) is order-homeomorphic to (V ↑ (X ), τ p, ≤), where τ p is the patch topology τ ∨ τ ∂ and ≤ is the inclusion order. Explain how the result from the preceding item, together with Theorem 4.41, show that upward Priestley compatible relations R ⊆ X × Y are in bijection with continuous order-preserving functions f : X → V (Y ).

Exercise 6.4.2 (a) Let 1 be the one-element space. Show that V ↑ (1) is homeomorphic to the Sierpinski space S. (b) Let X be a stably compact space. Show that V ↑ (X ) is homeomorphic to [X ∂, S] equipped with the compact-open topology. Exercise 6.4.3 Let X and Y be topological spaces. Recall that [X, Y ] denotes the set of continuous functions from X to Y . (a)

Give an example to show that the generating family {K → U | K ∈ K S(X ), U ∈ Ω(Y )}

(b)

may fail to be closed under finite intersection and union. Let x ∈ X, K, K1, K2 ∈ K S(X ), U, U1, U2 ∈ Ω(Y ), U ⊆ Ω(Y ) a directed family, and C ⊆ K S(X ) a filtering family. Prove: (1) (2) (3) (4) (5)

(c)

(d)

(K1 → U) ∩ (K2 → U) = (K1 ∪ K2 ) → U; (K → U1 ) ∩ (K → U2 ) = K → (U1 ∩ U2 ); (↑x → U1 ) ∪ (↑x → U2 ) = ↑x → (U1 ∪ U2 );

K → ( U ) = {K → W | W ∈ U }; and 

if X is a well-filtered space, then ( C) → U = {K → U | K ∈ C}.

Suppose that X is a locally compact space, that is, for any open set U and x ∈ U, there exists a compact set K such that x ∈ K ⊆ U. Prove that the specialization order on the space [X, Y ] coincides with the pointwise ordering. Suppose G ⊆ K S(X ) generates K S(X ) as a join-semilattice and that B is a base for Y which is closed under finite unions. Show that {K → U | K ∈ G, U ∈ B} is a subbase for the compact-open topology on [X, Y ].

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

226

6 Omega-Point Duality

Exercise 6.4.4 Let M be a distributive lattice and Y its Stone dual space. For every y ∈ Y , define η(y) := ↑y, which is a compact-saturated subset of Y , and thus an element of V ↑ (Y ). (a) (b) (c) (d)

Prove that η is a spectral embedding. Show that the map η is dual to the quotient of F (M) under the congruence given by the scheme (a ∨ b) ≈ a ∨ b for a, b ∈ M and ⊥ ≈ ⊥. Prove that, under the isomorphism of V ↑ (Y )  Filt(M) of Proposition 6.22, the image of η is mapped to the set of prime filters of Y . Let L be a distributive lattice with Stone dual space X. Using the previous items and Lemma 6.27, prove that a base for the (not necessarily spectral!) subspace [X, Y ] of [X, V ↑ (Y )] is given by the sets  a⇒ b := { f ∈ [X, Y ] | f [ a] ⊆  b}, for a ∈ L, b ∈ M.

Note. In the setting of Priestley spaces, the more delicate issue of a two-sided version of the Vietoris functor was studied in Palmigiano (2004). Restricted to Boolean spaces, the Vietoris functor was also extensively studied in the context of modal logic (Kupke et al., 2004). This analysis was recently extended to dualities for the category of compact Hausdorff spaces in Bezhanishvili et al. (2022). Notes for Chapter 6 The name “Stone space” has also been used for the more restricted class of spaces that we call Boolean spaces in this book. We prefer to avoid the name “Stone space” for a class of spaces, to avoid confusion. We do use the terminology “the Stone dual space of a lattice” to refer to the spectral space associated to a distributive lattice through Stone’s duality. Our choice of the terminology “spectral spaces” follows in particular the monograph Dickmann et al. (2019), the first chapters of which we recommend as useful complementary reading to the material in this chapter.

https://doi.org/10.1017/9781009349680.007 Published online by Cambridge University Press

7 Domain Theory

In this chapter, we develop some of the order-topological theory of dcpos and domains and provide the key duality-theoretic elements that were used in Abramsky (1991) to solve so-called domain equations. In denotational semantics, one seeks a category of mathematical objects, socalled denotational types, whose elements will model programs while the objects themselves model the “types” of the programs. In addition, program constructors should be modeled by functors, so-called type constructors. In particular, if X and Y are program types, then we want to be able to form the type which, given a program of type X as input, outputs a program of type Y . In the setting of a concrete category, this will mean that we want the set Hom(X, Y ) to be endowed with structure in a natural way that makes it an object of the category whenever X and Y are; a formalization of this idea leads to the categorical definition of a Cartesian closed category. 1 In this setting, expressions involving the type constructors correspond to formal specifications of a program’s properties. In particular, type equations of the form X  F (X ) correspond to recursive specifications, and their solutions are known as inductive types. The sought-after category should therefore be closed under a comprehensive set of type constructions, adequate for the needs of semantics of programming languages, as well as under solutions of type equations involving the type constructors. Finally, given the algorithmic nature of computing, it should also admit a reasonable notion of effective presentability. Scott and Strachey (1971) proposed to look for such a category within the category of dcpos with functions that are continuous with respect to the Scott topology, which we already briefly encountered back in Section 2.2 and will study in more detail starting in Section 7.1 of this chapter. Indeed, Scott originally considered the 1

We do not need to formally define this notion for our purposes in this chapter, the intuition of “Cartesian closed” meaning “has internal Hom-objects” suffices. We refer to, for example, p. 46 of Mac Lane and Moerdijk (1992) for the precise definition.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

228

7 Domain Theory

equation X  [X, X], which corresponds to the pure λ-calculus given by the program constructors of functional abstraction and composition. Scott’s solution for this equation looks within the category of dcpos with Scott continuous functions and seeks a dcpo X which is isomorphic to the dcpo of its Scott continuous endomorphisms. While the category of dcpos is Cartesian closed, the category as a whole fails to admit a reasonable notion of effective presentation and the general constructions get fairly wild. Domains, which we study in Section 7.1, were introduced precisely as a subcollection of dcpos with a notion of relatively finite presentability. Further restricting to algebraic domains gives actual finite presentability. Indeed, Scott’s solution to X  [X, X] is an algebraic domain. However, neither categories of domains or algebraic domains are Cartesian closed, so they do not provide quite the right setting either. A further property of Scott’s solution is that it is profinite and thus falls within Stone duality. Much further work by a large number of researchers confirmed that categorical methods, topology and in particular topological duality are central to the theory, see Scott and Strachey (1971), Plotkin (1976), Smyth and Plotkin (1982), Smyth (1983b), and Larsen and Winskel (1991). Abramsky went a step further in his view on duality-theoretic methods in this area. Rather than seeing Stone duality and its variants merely as useful technical tools for denotational semantics, he put Stone duality front and center stage. The centrality of duality and the general principles of the theory of domain equations is the subject of the two papers with the common title “domain theory in logical form”: while the main ideas are already put forward in Abramsky (1987), they come to their full and clear expression in Abramsky (1991), for which he won the IEEE Logic in Computer Science Test-of-Time Award in 2007. The subject is now commonly referred to by the title of these two papers or its acronym, DTLF. Abramsky casts Stone duality as the mechanism which links programs to their models. Thus, Abramsky defines a program logic, in which denotational types correspond to theories, and the ensuing Lindenbaum–Tarski algebras of the theories are bounded distributive lattices, whose dual spaces yield the domains as types. The constructors involved in domain equations have duals under Stone duality which are constructors in the program logic, and solutions are obtained as duals of the solutions of the corresponding equation on the lattice side. For this purpose, one needs a Cartesian closed category of domains that are spectral spaces in their Scott topology, and several such had in the meantime been discovered. We will study such spectral domains in Section 7.2, and the more general class of spectral dcpos. While the latter, larger class, is not directly used in domain theory in logical form, its characterization was only previously announced in a conference abstract (Erné,

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7 Domain Theory

229

2009), and for completeness’ sake we give a proof of it in Section 7.2, see Theorem 7.38. A reader who wants to get quickly to the domain theory in logical form material may safely skip that part of Section 7.2. In Abramsky (1987), duality is restricted to the so-called Scott domains. These domains are fairly simple and are closed under many constructors, including function space, sums, products, and upper and lower powerdomains, but they are not closed under the convex powerdomain (corresponding to the two-sided Vietoris construction in topology). In Abramsky (1991), it is shown that his program goes through for the larger category of bifinite domains, although the mathematics, and especially the duality theory, is much more involved. See also the book of Zhang (1991) where many of the technical aspects of this theory were developed concurrently, albeit without the central focus on Stone duality. We introduce bifinite domains and study them from a duality-theoretic perspective in Section 7.3. This category, which was originally introduced by Plotkin, has further closure properties. In particular, it is closed under the convex powerdomain construction. As shown in Smyth (1983a), this is optimal when restricted to countably based algebraic domains with least element. The PhD dissertation of Jung (1989) completely described all Cartesian closed categories of algebraic domains and, shortly thereafter, Jung introduced two new categories, of L- and FS-domains, respectively, and showed these to be the maximal Cartesian closed categories of continuous domains (Jung, 1990). This however is far from the end of the story as researchers in the area turned towards the probabilistic powerdomain, which forces one into the setting of continuous domains. In order to include the unit interval, which is central in probability theory, one has to go beyond the profinite setting. This leads to the problem of finding a Cartesian closed category of continuous domains which is closed under the probabilistic powerdomain. See Jung and Tix (1998) for more on this problem and Goubault-Larrecq (2019), which surveys what was known at that time. The notes at the end of this chapter give some more bibliographic details on specific points. We finish this extended introduction by giving a short outline of the chapter, in particular giving a road map for readers who want to get to Section 7.4, in some sense the heart of the chapter, as quickly as possible. Section 7.1 introduces the notion of a domain, which may be seen either as a special kind of poset or as a special kind of sober topological space. In Section 7.2, we study the intersection of the class of domains with the class of spectral spaces, and show that, under Stone duality, the domains in this intersection correspond to a very natural class of distributive lattices (Theorem 7.45). We also establish that these domains are exactly the ones satisfying the properties in the “2/3 SFP Theorem” of Plotkin, see Abramsky and Jung (1994, Proposition 4.2.17).

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

230

7 Domain Theory

Each of these first two sections also contains an excursion to a purely dualitytheoretic result; namely, in Section 7.1, we show that the Ω-pt duality of Chapter 6 restricts to the so-called Hoffmann–Lawson duality between completely distributive complete lattices and domains viewed as spaces (Theorem 7.21); in Section 7.2, in Theorem 7.38, we characterize the dcpos that are spectral spaces in their Scott topology, a theorem first announced in Erné (2009). These two results are not directly important for the domain-theoretic applications in the later two sections of the chapter and the rather intricate proofs may be skipped by a reader who wants to get to the duality-theoretic view on bifinite domains and domain equations as fast as possible. In Section 7.3, we introduce and study bifinite domains, which form the category of domains in which we will constructively solve domain equations, and which moreover are interesting objects for a duality theorist because of their self-dual nature. We finish the chapter in Section 7.4 by showing how the duality-theoretic analysis of the function space construction done in the previous chapter in Section 6.4 specializes in the case of bifinite domains, and how this allows one to reconstruct solutions to Scott’s equation X  [X, X]. 7.1 Domains and Hoffmann–Lawson Duality In a state space of computations, we may think of order as “further specification.” That is, two points p and q satisfy p ≤ q provided q is a further specified state of the computation than the state p. A bottom element may then be thought of as the state of no specification at all, and often domain theorists will consider posets with bottom (called pointed posets) as a convenient set-up. A top element, on the other hand, would further specify all computations in the space. Since one usually would want to consider states leading toward different computations, it is most common in domain theory to consider posets without a top. When a top is added, it is often thought of as the inconsistent, overspecified state. Further, since a complex program which computes its output by finite approximation may be seen as the directed join of the finite approximations, the posets considered in domain theory are often assumed to be closed under suprema of directed sets. The final ingredient is that any point should be obtainable by directed supremum from “finite” or at least “relatively finite” points below it. The pertinent mathematical notions abstracting these ideas are as follows. Definition 7.1 A partially ordered set P is directedly complete provided every directed subset of P has a supremum in P. A directedly complete partially ordered set is called a dcpo, for short. Let P and Q be dcpos and f : P → Q a map. Then, f is said to preserve directed joins provided f is order preserving and, for every directed set D, we have   f [D]. f D =

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

231

Note that since f is order preserving, f [D] is directed whenever D is (see Exercise 7.1.1) and thus, both suprema in this equation exist. We denote by dcpo the non-full subcategory of Pos consisting of dcpos with as morphisms those order-preserving maps which preserve directed joins. Remark 7.2 Here, we have included in the property of f preserving directed joins that f is order preserving, even though it is a consequence of preservation of directed joins. This reflects the fact that we consider dcpo as a subcategory of Pos. To avoid confusion, every time we state that a function preserves directed joins, we will make sure that it is also clear that this means that it must be order preserving. While this category is given entirely in order-theoretic terms, as we will see, it is actually isomorphic to the full subcategory of Top consisting of (the underlying sets of) dcpos equipped with the Scott topology of the dcpo. One significant and interesting feature of dcpos is that, viewing them as posets, they are first-order structures. This allows access to a class of topological spaces (non first-order) based on first-order structures. However, being closed under directed joins is not a first-order property. Recall from Section 2.2 that the Scott topology on a poset (P, ≤) is the topology σ(P, ≤) consisting of those sets U ⊆ P which are up-sets in the order and which are inaccessible by directed joins. That is, an up-set U ⊆ P is  Scott open provided whenever D ⊆ P is directed and D ∈ U then D ∩ U  ∅. Equivalently, we have that a set C ⊆ P is Scott closed if, and only if, C is a  down-set, and whenever D ⊆ P is directed and D ⊆ C then D ∈ C. That is, a down-set C is Scott closed if, and only if, C is closed under directed joins. The specialization order of the Scott topology on a poset P is equal to the original order on P (see Exercise 7.1.1(a)). A function f : P → Q between dcpos is said to be Scott continuous provided it is continuous with respect to the Scott topologies on P and Q. Remark 7.3 In this chapter, as in the modern literature on domain theory, open sets are up-sets. As also noted in Remark 3.22, this clashes somewhat with the choice in Priestley duality theory, which we adhered to up until Chapter 6 in this book, that represents a distributive lattice L as the clopen down-sets of its Priestley dual space X. We already encountered a similar clash in Chapter 6: Recall from Theorem 6.4 that the lattice of compact-opens of a spectral space (X, ρ) is isomorphic to the the lattice of clopen down-sets of (X, ρp, ≤), where ≤ is the opposite of the specialization order ≤ρ ; see also Figure 6.1. Note that another way of saying this is that if L is a distributive lattice with Priestley space (X, τ, ≤), then the associated spectral space (X, τ ↑ ) of open up-sets has L op as its lattice of compact-open sets. Using up-sets or down-sets to represent lattice elements is ultimately an arbitrary choice, and there are valid arguments for either choice. Being flexible about this type

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

232

7 Domain Theory

of “order yoga” is a somewhat cumbersome, but necessary, part of every duality theorist’s life. As a consequence, certain order-theoretic arguments in this chapter may sometimes look “upside down” from those in the earlier chapters of this book. We made the choice to adhere to the conventions from domain theory in this chapter and we will take care to warn the reader throughout the chapter when certain orders are the opposite of the orders used when we discussed Priestley duality. As a general remark, for later use in this chapter, we note already that, since Priestley duality is order reversing on morphisms, see Proposition 5.39, the duality we consider here is order preserving on morphisms. That is, if f , g : X ⇒ Y are two spectral maps between spectral spaces and K and L are the lattices of compact-opens of X and Y , respectively, with f ∗, g ∗ : L ⇒ K the dual lattice homomorphisms, then f ≤ g in the pointwise order on spectral maps with respect to the specialization order on Y , if, and only if, f ∗ ≤ g ∗ in the pointwise order on lattice homomorphisms with respect to the inclusion order on K. Proposition 7.4 Let P and Q be dcpos and f : P → Q a map. The following conditions are equivalent: (i) (ii)

f is order preserving and preserves directed joins; and f is Scott continuous.

Proof To show that (i) implies (ii), suppose f is order preserving and preserves directed joins and let U ⊆ Q be Scott open. First, since f is order preserving,  f −1 (U) is an up-set. Now, suppose D ⊆ P is directed and D ∈ f −1 (U). Then, f [D] is directed in Q by Exercise 7.1.1(b) and, as f preserves directed joins, it   follows that f [D] = f ( D) ∈ U. Since U is Scott open, it follows that there is d ∈ D with f (d) ∈ U and thus d ∈ f −1 (U) and we have proved that f −1 (U) is Scott open. For the converse, suppose f is Scott continuous. By Exercise 2.2.5 in Chapter 2, any continuous function is order preserving with respect to the specialization order. Now, the specialization order of the Scott topology of a dcpo is equal to the order of the dcpo (see Exercise 7.1.1(a)). So, f is order preserving for the original orders on P and Q, respectively. Now, suppose D ⊆ P is directed. Since f is order preserving, f [D] is directed  and thus f [D] exists in Q. Also since f is order preserving, it follows that    f [D] ≤ f ( D). Now, consider C = ↓( f [D]). Since it is a principal downset, it is closed in the Scott topology. By (ii), f −1 (C) is closed in the Scott topology on P. Also note that by the definition of C and as f is order preserving we have   D ⊆ f −1 (C). Therefore, D ∈ f −1 (C) or, equivalently, f ( D) ∈ C. That is,   f ( D) ≤ f [D], as required. 

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

233

Corollary 7.5 The category dcpo is isomorphic to a full subcategory of the category Top of topological spaces with continuous maps. Domain theorists often implicitly and harmlessly switch between the two perspectives that are provided by Corollary 7.5. However, this practice, combined with the practice of suppressing the structure in the denotation of a mathematical structure may lead to complications: There are topological spaces X which are dcpos in their specialization order, even though the original topology on X is not the Scott topology. To be able to state our results in this section, it will sometimes be convenient to keep the distinction between the two perspectives clear; to do so, when (P, ≤ P ) is a dcpo, we refer to (P, σ(P, ≤ P )) as the space of the dcpo P. We also note that, despite the isomorphism of categories in Corollary 7.5, unexpected things can happen in switching perspectives if one is not careful; for example, the Scott topology on the Cartesian product of two dcpos with the pointwise ordering need not be equal to the product topology of the Scott topologies on the two dcpos, see Gierz et al. (2003, Exercise II-4.26). From a categorical point of view, this means that, if X and Y are dcpos, then X × Y with the pointwise order is the product of X and Y in the ambient category of posets, but the space of the dcpo X × Y is not necessarily the product in the ambient category of topological spaces. When we restrict to the full subcategory of dcpo on the objects that are domains, see Definition 7.10, this problem does not occur and finite products of Scott topologies equal the Scott topology of the topological products, see Gierz et al. (2003, Corollary II-4.14). The notions of dcpo and of Scott continuity also arise naturally from a purely topological point of view, or more specifically, from the point of view of the Ω-Pt duality, as follows. Proposition 7.6 If (X, τ) is a sober topological space, then (X, ≤τ ) is a dcpo and τ ⊆ σ((X, ≤τ )). Thus, any sober topological space X comes with a naturally associated dcpo, that we call the dcpo of the sober space X. The above proposition then says that the Scott topology of the dcpo of a sober space is always finer than (i.e., has at least as many open sets as) the original sober topology itself. Before giving a proof of this proposition, we note that a proof via Ω-Pt duality was outlined for spaces of the form Pt(L) where L is any frame in Exercise 6.3.6. Essentially, it is a direct consequence of the fact that the specialization order of Pt(L) is the inclusion order on completely prime filters and that these are closed under directed unions. By the Ω-Pt duality, the sober spaces are, up to homeomorphism, precisely the spaces of the form Pt(L). Thus, this proves the proposition. Here we give a direct proof, not invoking the duality.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

234

7 Domain Theory

Proof of Proposition 7.6 Suppose (X, τ) is sober. It suffices to prove the following: Claim For any closed F ⊆ X and for any D ⊆ F directed in the specialization  order ≤τ , the supremum D exists and belongs to F. Indeed, this claim implies the proposition, for the following reasons. First of all, if D is an arbitrary directed subset of X, then choosing F = X we just get that  D exists. Further, since a subset of a poset is Scott closed if, and only if, it is a down-set and is closed under directed joins, we will in fact have shown that every τ-closed set (which is necessarily a down-set in the specialization order) is Scott closed. Thus, τ is contained in the Scott topology as claimed. We now prove the claim. Let F be closed and D ⊆ F directed in ≤τ . Define F = {U ∈ τ | U ∩ D  ∅}. It is easy to check that F is a completely prime filter; we only show that F is closed under binary intersection and leave the other parts to the reader. If U, V ∈ F , then there are p, q ∈ D with p ∈ U and q ∈ V . Since D is directed, there is some r ∈ D which is above both p and q. Now, since open sets are always up-sets in the specialization order, it follows that r ∈ U ∩ V and thus U ∩ V ∈ F thus showing that F is closed under binary intersection. Now, since F is a completely prime filter and X is sober, it follows from the definition of sobriety (Definition 6.18) that there exists x ∈ X so that F = N (x). Since every open neighborhood of x intersects D ⊆ F and thus F and since F is closed, it follows that x ∈ F.  Finally, we show that D = x. Let y ∈ D. We need to show that y ≤τ x. Let U ∈ τ with y ∈ U. Then, D ∩ U  ∅ and thus U ∈ F = N (x). That is, x ∈ U and thus we have shown that every open containing y contains x. That is, y ≤τ x as required. On the other hand, suppose z ∈ X is an upper bound for D. We show that x ≤τ z. Let U ∈ τ with x ∈ U, then U ∈ N (x) = F and thus D ∩ U  ∅. Pick y ∈ D ∩ U. Since z is an upper bound of D, y ≤τ z and thus z ∈ U. That is, x ≤τ z as required.  Remark 7.7 The space of a dcpo does not need to be sober, see Johnstone (1981). In fact, understanding the frames of Scott open sets of dcpos seems a difficult problem, see Ho et al. (2018), in which it is shown that dcpos are not determined up to isomorphism by their closed set lattices (which are of course isomorphic to the order dual of the frames of Scott open sets). This will not be a problem for us here, because we will soon restrict to a class of dcpos all of whose spaces are sober, see Proposition 7.19. Example 7.8 Let X be any set. Then, (X, =) is a dcpo and σ(X, =) = α(X, =) = δ(X, =) = P (X ).

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

235

That is, relative to the trivial order, the Scott topology is equal to the Alexandrov topology and these are equal to the discrete topology on X. In particular, this example shows that the space of the dcpo of any T1 topological space is discrete. Thus, the inclusion in Proposition 7.6 may very well be strict. Example 7.9 The unit interval in its usual order ([0, 1], ≤) is a complete lattice, so in particular a dcpo. Its Scott topology is generated by the half-open intervals, that is, the sets of the form (a, 1] = {x ∈ [0, 1] | a < x}, for a ∈ [0, 1], and is thus equal to the upper topology; the resulting topological space is in fact a stably compact space (see Section 2.3). The reader may verify that the corresponding compact ordered space carries the usual compact Hausdorff topology of the unit interval inherited from the usual topology on the real line and the order is the usual order inherited from the reals. The following notions are of fundamental importance in computer science applications of dcpos.



Definition 7.10 An element k in a dcpo X is said to be a compact element provided  for all directed D ⊆ X with k ≤ D there is d ∈ D with k ≤ d. We will denote the set of compact elements of X by K(X ). Compact elements are sometimes called finite elements in the literature. Let x, y ∈ X. We say that x is way below y and write x ) y provided that, for all  directed D ⊆ X, if y ≤ D, then there is d ∈ D with x ≤ d. Further, we denote by y the set of all elements of X that are way below y. That is,





y = {x ∈ X | x ) y}



and y is defined similarly. We call a dcpo X a domain, also known as a continuous dcpo, provided each element of X is the directed join of the elements way below it. More explicitly, for a dcpo X to be a domain, for any x ∈ X, the set x must be directed, and its supremum must be x. We call X an algebraic dcpo or an algebraic domain provided each element of X is the directed join of the compact elements below it.



As the nomenclature “way below” suggests, x ) y implies x ≤ y for any elements x, y of a dcpo X. Also, an element x in a dcpo X is compact if, and only if, x ) x (see Exercise 7.1.3). We note that, in the definition of a domain, it suffices to assume that each element x of the dcpo is the directed join of some directed subset of the elements way below x; every element of the domain will then in fact be equal to the directed join of the set x (see Exercise 7.1.5). This fact can be quite useful in proofs, as we will see, for example, in the proof of Proposition 7.31. We highlight an alternative characterization of algebraic domains that will be particularly important throughout the chapter, starting from Section 7.2. For a poset

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

236

7 Domain Theory

P, we call an order ideal of P a down-set that is directed; note that this definition generalizes the notion of ideal for a lattice L, see also the remarks following Theorem 6.19. We denote by Idl(P) the collection of ideals of P, ordered by inclusion. Now, Idl(P) is always an algebraic domain and an algebraic domain X is always isomorphic to Idl(K(X )); thus, a domain X is algebraic if, and only if, it is isomorphic to Idl(K(X )). Exercise 7.1.11 asks you to prove this equivalence, via some other equivalent characterizations of algebraic domains. Example 7.11 For any set X, all elements of the dcpo (X, =) are compact and thus it is an algebraic domain. Example 7.12 If (P, ≤) is any finite poset, then it is an algebraic domain; indeed, any directed subset of P is also finite, and therefore contains a maximum element. From this, it follows that x)y if, and only if, x ≤ y, so all elements are compact. Example 7.13 In ([0, 1], ≤), we have x)y if, and only if, x = 0 or x < y. It follows that 0 is the only compact element and that the unit interval is a domain which is not algebraic. Example 7.14 Let X be a set. The partial order (P (X ), ⊆) is an algebraic domain: We have x)y if, and only if, x ⊆ y and x is finite (see Exercise 7.1.4). Thus, all finite subsets are compact elements of P (X ) and each subset is the directed union of its finite subsets. Example 7.15 Let (X, τ) be a topological space. An open subset K ⊆ X is compact in the topological sense if, and only if, K is a compact element of the frame ΩX. That is, K(Ω(X )) is equal to the set of compact-open subsets of X, which was denoted by K O(X ) in Chapter 6. Example 7.16 Let Part(X ) denote the poset of partial functions on a set X with the order given by f ≤ g if, and only if, g extends f . That is, dom( f ) ⊆ dom(g) and f (x) = g(x) for all x ∈ dom( f ). Then again, the finite partial functions (i.e., those with finite domain) are the compact elements and Part(X ) is an algebraic domain which is not a lattice (see Exercise 7.1.4). It does however have the property that all principal down-sets are complete lattices; in fact, they are complete and atomic Boolean algebras. Note that the total functions on X are the maximal elements of Part(X ). Toward proving that the space of a domain is always sober (Proposition 7.19), we now examine the way below relation on a domain in more detail. Definition 7.17 A binary relation R on a set X is called interpolating if R ⊆ R · R. Note that a relation R is idempotent, that is, R· R = R if, and only if, R is transitive and interpolating.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

237





Lemma 7.18 Let X be a domain. Then, ) is transitive, interpolating and, for each x ∈ X, the set x is Scott open. Furthermore, U ⊆ X is Scott open if, and only if,  U= x. x ∈U



Proof Note that ) is transitive on any dcpo (see Exercise 7.1.3). To see that ) is interpolating, let x, y ∈ X be such that x ) y. Since X is continuous, we have that x is way below   y= {z | z ∈ X, z)y} y=   = z | z ∈ X, z)y  = Z,

where Z := { z | z ∈ X, z ) y}. Since y is directed, the collection { z | z ∈ X, z ) y} is a directed family of sets in the inclusion order. Combining this with the fact that each z for z ) y is directed, it follows that the set Z is directed. Thus, there exist z, z  ∈ X with x ≤ z  ) z ) y. It follows that x() · ))y (see Exercise 7.1.3).  To show that x is Scott open, let D ⊆ X be directed with x) D. Then, there   is x  ∈ X with x ) x  ) D. Since x  ) D, there is d ∈ D with x  ≤ d. Now, since x ) x  it follows that x ) d and thus x ∩ D  ∅, showing that x is Scott

open. Consequently, any set U satisfying U = x ∈ U x is also Scott open. Finally, let U ⊆ X be Scott open. Clearly, since U is an up-set, in particular

x ∈U x ⊆ U. For the reverse inclusion, let y ∈ U. As X is a domain, is the directed join of the set y. Since U is Scott open it follows that there is x ) y with

 x ∈ U. Thus, y ∈ x ∈U x, as required. 























Proposition 7.19 Any domain is sober in its Scott topology. Proof Let X be a domain and suppose F ⊆ X is a join-irreducible closed set for the Scott topology. We need to show that F = ↓x for some x ∈ X. Consider the set 

F = {y ∈ X | ∃x ∈ F y)x} 

= {y ∈ X | y ∩ F  ∅}. 



We will first show that F is directed, so that it has a supremum, x, and we will then show that F = ↓x. Note first that F, being join irreducible, is non-empty. Pick x ∈ F. Then, as X is a domain, x is directed and in particular non-empty. Thus, F is not empty. Let y1, y2 ∈ F be arbitrary. Define Fi = ( yi ) c for i = 1, 2. Then, since yi ∈ F, there is x i ∈ F with yi ) x i and thus F  Fi for both i = 1, 2. Now since F1 and F2 are closed in the Scott topology by Lemma 7.18 and since F is a join-irreducible closed set, it follows that 







https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press





F  F1 ∪ F2 = ( y1 ∩ y2 ) c .

238

7 Domain Theory



















Thus, pick z ∈ F ∩ y1 ∩ y2 . Since yi is open for both i = 1, 2, it follows that  y1 ∩ y2 is open. Therefore, since z = z ∈ y1 ∩ y2 , there exists z  )z with  z ∈ y1 ∩ y2 . Thus, in particular, y1 ≤ z  and y2 ≤ z . Also, since z  )z ∈ F it  F. We follows that z  ∈ F and we have shown that F is directed. Let x = show that F = ↓x. First, since F ⊆ ↓F ⊆ F and since F is closed, it follows that  F ∈ F. Since F is a down-set, we get the containment ↓x ⊆ F. Conversely, x= for any x  ∈ F, we have x  ⊆ F and thus   x = x ≤ F = x. 

















That is, F ⊆ ↓x as required.



Hoffmann–Lawson Duality As a consequence of Corollary 7.5 and Proposition 7.19, the category Domain of domains with order-preserving maps that preserve directed joins (or equivalently, Scott continuous functions) is (isomorphic to) a full subcategory of the category Sober of sober topological spaces. Thus, it is natural to ask which category of spatial frames is dual in the Ω-Pt duality to the category Domain. The appropriate frames are the completely distributive ones and the resulting duality is the Hoffmann– Lawson duality, as we will prove now. As noted in the introduction to this chapter, while this is an interesting excursion to a classical result in domain theory, it is not directly needed for the applications in Section 7.4, and can be skipped by readers wanting to get to those applications as quickly as possible. Definition 7.20 A complete lattice L is said to be completely distributive provided, for any family { Ai }i ∈I of subsets of L, we have     im(Φ) | Φ : I → L such that Φ(i) ∈ Ai for each i ∈ I . Ai = i ∈I

For a family { Ai }i ∈I of sets, the functions Φ : I → each i ∈ I are called choice functions on { Ai }i ∈I .

i ∈I

Ai such that Φ(i) ∈ Ai for

Since all suprema exist in a frame, any frame is in fact a complete lattice (see Exercise 1.2.2 and Exercise 6.2.2). Note that a completely distributive complete lattice is in particular a frame (Exercise 7.1.8), but not conversely (Exercise 7.1.6(b)). We will show that Ω-Pt duality further restricts to a duality between domains and completely distributive complete lattices. Theorem 7.21 The Ω-Pt duality between spatial frames and sober spaces cuts down to a duality between the category CDFrame of completely distributive complete lattices with frame homomorphisms and the category of domains, Domain.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

239

Ω CDFrameop

Domain. Pt

The remainder of this section is dedicated to proving Theorem 7.21. We first give an outline of the proof. Outline of proof of Theorem 7.21 We have that: (a) (b) (c) (d)

If X is a domain, then X is sober (Proposition 7.19). If X is a domain, then Ω(X ) is completely distributive (Proposition 7.27). If L is completely distributive, then L is spatial (Corollary 7.30). If L is completely distributive, then Pt(L) is a domain and its topology is the Scott topology (Proposition 7.31 and Lemma 7.33).

This is precisely what is needed to show that the Ω-Pt duality cuts down to a duality between domains and completely distributive complete lattices.  Toward proving Proposition 7.27, we begin by studying in slightly more detail the class of completely distributive complete lattices. First, we give a simpler description of complete distributivity. Proposition 7.22 A complete lattice L is completely distributive if, and only if, for all families (Di )i ∈I of down-sets of L, we have   Di  . i ∈I  i ∈I      Proof First note that for any family of sets ( Ai )i ∈I , i ∈I Ai = i ∈I ↓Ai and,

for any choice function Φ : I → i ∈I ↓Ai , there is a choice function Φ : I →  

 im(Φ) ≤ im(Φ ). It i ∈I Ai with Φ(i) ≤ Φ (i) for every i ∈ I, and thus with follows that it suffices to consider families of down-sets in the definition of complete distributivity. Now the proposition follows if we can show that, for any family {Di }i ∈I of down-sets of L, we have   (7.1) Di = im(Φ) | Φ is a choice function for {Di }i ∈I . 



Di =

i ∈I

Let a ∈ i ∈I Di . Then, the constant function Φa on I given by Φ(i) = a for all i ∈ I  is a choice function for {Di }i ∈I and im(Φa ) = a. This proves the left-to-right containment in (7.1).

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

240

7 Domain Theory

On the other hand, if Φ is any choice function for {Di }i ∈I , then for any i ∈ I, the    element im(Φ) is below Φ(i) ∈ Di , and is thus in i ∈I Di . In order to understand completely distributive complete lattices, we introduce a strengthened variant of the way below relation, which is a relativized version of complete join-primeness in the same way that the way below relation is a relativized version of compactness for elements of a dcpo.  

Definition 7.23 Let L be a complete lattice and a, b ∈ L. We write b ≪ a, or  equivalently b ∈ a, provided that, for any subset S of L, if a ≤ S, then there exists s ∈ S such that b ≤ s. An element a in a complete lattice is called completely join prime if a ≪ a (see Exercise 1.3.10). Note that a ≪ b clearly implies a ) b. The following theorem is due to Raney (1953).

 

Theorem 7.24 (Raney’s theorem) A complete lattice L is completely distributive if, and only if,  a. (7.2) for every a ∈ L, a=

 

Proof Suppose (7.2) holds and let {Di }i ∈I be a collection of down-sets of L. We define    Di  . Di, d  := d := i ∈I  i ∈I    We want to show that d = d . In fact, for each i ∈ I, since i ∈I Di ⊆ Di we have  d  ≤ Di , so d  ≤ d. We now show the reverse inequality, d ≤ d . To this end, let  b ≪ d. Then, b ≪ Di for each i ∈ I. By the definition of ≪ there is, for each i ∈ I, an element d i ∈ Di with b ≤ d i . Since each Di is a down-set, it follows that  b ∈ Di for each i ∈ I and thus b ∈ i ∈I Di . That is, d⊆ Di i ∈I

 

 and thus by (7.2) we have the desired inequality d = d ≤ d . For the converse implication, suppose L is completely distributive. Note that, for any a ∈ L, we have     a= S | S ∈ D(L) and a ≤ S . This is because a is clearly a lower bound of the collection we are taking the  infimum of and S := ↓ a is a down-set of L with a = S. Now, applying complete distributivity we obtain    

a= S | S ∈ D(L) and a ≤ S .

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

Finally, observe that

S | S ∈ D(L) and a ≤

  S = a.



 



241

Corollary 7.25 Any completely distributive complete lattice is continuous. This is a consequence of the relation between ) and ≪ (see Exercise 7.1.9). Corollary 7.26 The relation ≪ is interpolating on a completely distributive complete lattice. The proof of Corollary 7.26, which we ask you to give in Exercise 7.1.9, is a simpler version of the corresponding fact for ) on a domain (Lemma 7.18). For a more substantial consequence of Raney’s result, we prove one direction of Hoffmann–Lawson duality. Proposition 7.27 Let X be a domain. The frame σ(X ) of Scott open subsets of X is completely distributive. 









Proof Let U ∈ σ(X ). Recall Lemma 7.18, in which we showed that, if X is a

domain then, for each x ∈ X, the set x is Scott open and U = x ∈U x. Thus, by Theorem 7.24, we may conclude that σ(X ) is completely distributive if we can

show that x≪U for each x ∈ U. To this end, note that if x ∈ U and U = i ∈I Ui ,  then there is i ∈ I with x ∈ Ui and thus x ⊆ ↑x ⊆ Ui and indeed x≪U. To prove that, conversely, any completely distributive complete lattice L is isomorphic to one of the form σ(X ) for some domain X, we need a few lemmas. Lemma 7.28 Let L be a frame and let F, G be completely prime filters in L. If there is a ∈ L with F ⊆ ↑a ⊆ G, then F)G.

Proof If {Fi }i ∈I is a directed family of completely prime filters and G = i ∈I Fi ,  then a ∈ Fi for some i ∈ I and thus F ⊆ ↑a ⊆ Fi . Lemma 7.29 Let L be a completely distributive complete lattice and a, b ∈ L, then a≪b if, and only if, there exists a completely prime filter F in L such that ↑b ⊆ F ⊆ ↑a. Proof Suppose a≪b. Since ≪ is interpolating by Corollary 7.26, there is a sequence {bn }n∈N with a≪bn+1 ≪bn ≪b for all n ∈ N. One may verify that F =

n∈N ↑bn is a completely prime filter (see Exercise 7.1.10) and that ↑b ⊆ F ⊆ ↑a. For the converse, notice that if ↑b ⊆ F ⊆ ↑a for some completely prime filter F  in L and b ≤ S, then there is s ∈ S with s ∈ F and thus a ≤ s. That is, a ≪ b.  Corollary 7.30 Any completely distributive complete lattice is a spatial frame. Proof Let L be a completely distributive complete lattice. If a, b ∈ L with a  b then, by Raney’s Theorem, there is c ∈ L with c ≪ a but not c ≪ b. By Lemma 7.29 pick a completely prime filter F of L with ↑a ⊆ F ⊆ ↑c. By Lemma 7.29 and

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

242

7 Domain Theory

the fact that not c≪b, we must have ↑b  F. That is, a ∈ F and b  F and thus L is spatial.  Recall from Definition 6.12 that we adopt the “neutral space” notation for Pt(L), analogously to what we did in Notation 3.20 for the Priestley dual space of a distributive lattice. That is, we consider the set underlying Pt(L) as a fresh set of “names” that is in bijection with the set of completely prime filters of L. We denote elements of Pt(L) by x, y, z, . . ., and the corresponding completely prime filters respectively by Fx, Fy, Fz , et cetera.

 

Proposition 7.31 If L is a completely distributive complete lattice then the dcpo Pt(L) is a domain.  b = b ∈ Fy , there is a ∈ Fy with Proof Let y ∈ Pt(L) and b ∈ Fy . Since a≪b. By Lemma 7.29, it follows that there is x b ∈ Pt(L) with ↑b ⊆ Fxb ⊆ ↑a. Therefore, Fxb ⊆ ↑a ⊆ Fy and, by Lemma 7.28, we have x b )y. Now, notice that   {↑b | b ∈ Fy } ⊆ {Fxb | b ∈ Fy } Fy =  ⊆ {Fx | ∃a ∈ L Fx ⊆ ↑a ⊆ Fy } ⊆ Fy . Consequently, y=



{x ∈ Pt(L) | ∃a ∈ L

Fx ⊆ ↑a ⊆ Fy }. 

 

By Lemma 7.28, it follows that y is the join of a subfamily of y. If we can show that this collection is directed, we can conclude that Pt(L) is a domain by Exercise 7.1.5. To this end suppose Fx ⊆ ↑a ⊆ Fy and Fz ⊆ ↑b ⊆ Fy . Since both a and b  c, there is d ≪ c with belong to Fy , we have c = a ∧ b ∈ Fy . Now, since c = d ∈ Fy . By Lemma 7.29, it follows that there is s ∈ Pt(L) with ↑c ⊆ Fs ⊆ ↑d. Since a and b are both above c, it follows that Fx ⊆ ↑a ⊆ ↑c ⊆ Fs and Fz ⊆ ↑b ⊆ ↑c ⊆ Fs . Finally, since d ∈ Fy , we have Fs ⊆ ↑d ⊆ Fy and thus x, z ) s ) y. We conclude  that {x ∈ Pt(L) | ∃a ∈ L Fx ⊆ ↑a ⊆ Fy } is indeed directed. Corollary 7.32 If L is a completely distributive complete lattice and x, y ∈ Pt(L), then x)y if, and only if, there exists an a ∈ L with Fx ⊆ ↑a ⊆ Fy . Proof By Lemma 7.28 the “if” part is always true. For the converse, note that, in the proof of Proposition 7.31, we proved that if y ∈ Pt(L), then  y= {z ∈ Pt(L) | ∃a ∈ L Fz ⊆ ↑a ⊆ Fy } and that this join is directed. Thus, if x ) y, then there is z ∈ Pt(L) and a ∈ L so  that Fz ⊆ ↑a ⊆ Fy and x ≤ z, but then Fx ⊆ ↑a ⊆ Fy . Lemma 7.33 If L is a completely distributive complete lattice, then the topology on Pt(L) is equal to the Scott topology of its specialization order.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

243

Proof Since any space in the image of the Pt functor is sober, it follows by Proposition 7.6 that the topology of Pt(L) is contained in the Scott topology. For the converse, let U ⊆ Pt(L) be Scott open. Recall from Definition 6.12 that the topology of Pt(L) consists of the sets  a = {x ∈ Pt(L) | a ∈ Fx } for a ∈ L. Let   a= Fx | x ∈ U . We claim that U =  a. First, if a ∈ Fy , then as Fy is completely prime, there is x ∈ U  with Fx ∈ Fy . It follows that Fx ⊆ Fy or equivalently that x ≤ y and thus y ∈ U. That is,  a ⊆ U. On the other hand, if x ∈ U then, since Pt(L) is a domain, there is y ∈ U with y ) x. By Corollary 7.32, there is b ∈ L with Fy ⊆ ↑b ⊆ Fx . It follows  a.  that b ≤ Fy ≤ a. Finally, since b ∈ Fx also a ∈ Fx and x ∈  This concludes the last piece of the proof of Theorem 7.21. Exercises for Section 7.1 Exercise 7.1.1 Let P and Q be dcpos and f : P → Q a map. (a) (b) (c)

Show that the specialization order of the topological space (P, σ(P, ≤)) is the original order ≤ on P. Show that if f is order preserving and D ⊆ P is directed, then so is f (D). Show that S : dcpo → Top (P, ≤) → (P, σ(P, ≤)) f → f is a functor whose image is a full subcategory of Top. Further, show that if S(P) and S(Q) are homeomorphic as topological spaces, then P and Q are isomorphic as posets.

Exercise 7.1.2 Let (X, ≤) be a partially ordered set. (a) (b) (c) (d)

Show that the Scott topology on the poset D = D(X, ≤), ordered by subset inclusion, is equal to its upper topology. Show that these topologies are generated by the principal up-sets Ux = {V ∈ D | x ∈ V }, where x ranges over the elements of X. Show that the resulting space is stably compact. Show that the associated compact ordered space is a Priestley space.

Exercise 7.1.3 Let (X, ≤) be a dcpo and x ∈ X. (a)

Show that x is way below itself if, and only if, x is a compact element of X.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

244

Show that the way below relation is contained in the order relation. Show that ≤ · ) · ≤= ). Show that the set x is closed under any existing binary joins. That is, if y, z ∈ x and y ∨ z exists in X, then y ∨ z ∈ x. 

(b) (c) (d)

7 Domain Theory





Exercise 7.1.4 Let X be a set. (a) (b) (c) (d) (e)

Prove that in the dcpo (P (X ), ⊆), x)y if, and only if, x ⊆ y and x is finite. Draw the Hasse diagram of the partial order Part(X ) in case X has two elements. Characterize the way below relation in the dcpo (Part(X ), ≤) of Example 7.16. Show that (Part(X ), ≤) is not a lattice when X has at least two elements. Show that, for any f ∈ Part(X ), the sub-poset ↓ f is isomorphic to (P (dom( f )), ⊆).

Exercise 7.1.5 Prove that if X is a dcpo in which every element is the join of some directed set of elements way below it, then X is a domain. Exercise 7.1.6 Prove each of the following statements: (a) (b) (c) (d)

A complete lattice L is completely distributive if, and only if, its order dual is completely distributive. There are frames which are not completely distributive. There are completely distributive frames with no completely join-irreducible or completely meet-irreducible elements. A complete Boolean algebra is completely distributive if, and only if, it is atomic (see Exercise 6.3.2).

Exercise 7.1.7 An element j of a complete lattice L is called completely join irreducible if, for any S ⊆ L, if j = S, then j ∈ S. (a) (b)

Show that a completely join-prime element is always completely join irreducible. Show that, if L is a frame, then any completely join-irreducible element is completely join prime.

Exercise 7.1.8 This exercise compares the complete distributivity law to the strictly weaker frame distributivity law. (a)

Prove that a complete lattice L is a frame if, and only if, the complete distributive law holds in L for finite collections, that is, if for any finite index set I and any collection { Ai }i ∈I of subsets of L,    im(Φ) | Φ : I → L a choice function for { Ai }i ∈I . Ai = i ∈I

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.1 Domains and Hoffmann–Lawson Duality

(b)

245

Formulate a “directed distributive law” (DDL) such that a complete lattice is completely distributive if, and only if, it is a frame that satisfies DDL.

Exercise 7.1.9 Let L be a completely distributive complete lattice and a ∈ L. 

 

(a) (b) (c)

Show that any finite join of a subset of a is in a. Hint. Use Exercise 7.1.3(d). Show that L is continuous. Show that ≪ is interpolating. Hint. The proof is similar to the corresponding proof for ) given in Lemma 7.18.

Exercise 7.1.10 Let L be a frame and S ⊆ L which is filtering with respect to ≪. That is, if a, b ∈ S, then there is c ∈ S with c ≪ a and c ≪ b. Show that F = ↑S is a completely prime filter of L. Exercise 7.1.11 Let X be a domain. Recall that K(X ) denotes the poset of compact elements of X, and for a poset P, Idl(P) denotes the collection of order ideals of P; that is, down-sets that are up-directed. Show that the following conditions on X are equivalent: (i) (ii) (iii) (iv) (v)

X is algebraic; X  Idl(K(X )); the frame of opens of X is isomorphic to U (K(X )); the frame of opens of X is isomorphic to U (P) for some poset P; and X is isomorphic to Idl(P) for some poset P.

Exercise 7.1.12 Let P and Q be posets. A relation R ⊆ P × Q is called approximable provided ≥ P · R · ≥Q = R and R[p] = {q ∈ G | pRq} is directed for each p ∈ P. Show that the category Alg of algebraic domains is equivalent to the category Posapprox of posets with approximable relations via the functors which send an algebraic domain to its poset of compact elements and a poset to the algebraic domain of its order ideals. Note. Exercise 6.3.7 is related to this exercise in the case where the Scott topology on the algebraic domain is spectral, a situation that we will study in Section 7.2. The notion of bases and approximable relations between them allows one to generalize the equivalence between Posapprox and algebraic domains in this exercise to continuous domains in general, see Abramsky and Jung (1994, Section 2.2.6). Exercise 7.1.13 Let X be an algebraic domain. Show that U ⊆ X is compact-open in X if, and only if, U = ↑F for some finite F ⊆ K(X ). Conclude that the compactopen subsets of an algebraic domain form a base for the Scott topology on X which is closed under finite unions.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

246

7 Domain Theory

7.2 Dcpos and Domains That Are Spectral It is particularly interesting to know which dcpos and domains are such that their associated space is spectral. The topological spaces that are simultaneously spectral and Scott topologies of their specialization order have been characterized by Marcel Erné. While these spaces originally went by the name hyperspectral spaces, we will call them spectral dcpos here, in light of our convention that a dcpo is always equipped with its Scott topology. The first result of this section, Theorem 7.38, characterizes spectral dcpos as the coherent sober spaces which have a base of open finitely generated up-sets. While this result is not immediately needed for domain theory in logical form, we believe it answers a very natural question about the relationship between spectral spaces and dcpos. We therefore include a full proof of it here, as far as we are aware for the first time in writing, since the statement was so far only available in a conference abstract (Erné, 2009). Again, a reader who wants to get to the applications of duality to domain theory in logical form as quickly as possible may just read the relevant definitions and statement of Theorem 7.38. An important class of spaces that is directly relevant to the rest of this chapter and that we begin to study on p. 251 is obtained by restricting the class of spectral dcpos to those that are also domains; we refer to these objects as spectral domains here. These spaces have very nice descriptions both as spaces and as posets, which we will study in detail in the second part of this section. We will show in Theorem 7.47 that, in topological terms, these domains are precisely the spectral spaces whose compact-opens are finite unions of union-irreducible compact-opens, and in order-theoretic terms they are precisely the completions under directed joins of so-called finitely MUB-complete posets, see Corollary 7.48. In Section 7.3, we will introduce a further subcategory of the category of spectral domains, namely the bifinite domains. These are central to domain theory in logical form as treated in Abramsky (1991). Figure 7.1 gives an overview of the relevant classes of spaces and their relationships.

Spectral Dcpos Our goal in this subsection is to characterize spectral dcpos, defined as follows. Definition 7.34 A topological space (X, τ) is called a spectral dcpo if the topology τ is spectral and τ is equal to the Scott topology of its specialization order. The characterization will make crucial use of the following notion of a finitely generated up-set.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.2 Dcpos and Domains That Are Spectral

247

Top dcpo

Sober

Stably compact Domain Spec Bif

Figure 7.1 The various spaces of interest: Within the sober spaces, stably compact spaces allow a duality between open and compact-saturated sets, and Spec is a special case of this (see Chapter 6). Within the dcpos, Domain is a particularly important class of spaces allowing “relatively-finite” approximation. We will identify both the spaces in the intersection of dcpo and Spec, namely spectral dcpos, and the spaces in the intersection of Domain and Spec, namely spectral domains. Bifinite domains form a Cartesian closed category, Bif, of spaces in the intersection of Domain and Spec that is the setting of DTLF as elaborated by Abramsky.

Definition 7.35 Let X be a poset. A finitely generated up-set is an up-set T ⊆ X such that T = ↑F for some finite subset F ⊆ X. In a topological space X, a finitely generated up-set in the specialization order of X is sometimes called a hypercompact saturated subset or a tooth of X, but we will also use the term finitely generated up-set for this concept, where it is understood that we mean finitely generated up-set in the specialization order. A finitely generated up-set in a topological space X is clearly compact, since only finitely many opens are needed to cover a finite set and, since opens are up-sets, they cover the up-set generated by any set they contain. Importantly for us, in Scott topologies, the converse also holds. Lemma 7.36 Let X be a poset and V ⊆ X a Scott open subset. Then, V is compact in the Scott topology on X if, and only if, V is a finitely generated up-set.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

248

7 Domain Theory

Proof We only need to show the necessity, as sufficiency holds in any topological space. Any compact-saturated subset of a T0 space is the up-set of its minimal points (see Exercise 7.2.1). We show that, for each x ∈ min(V ), the set Ux = (V − min(V )) ∪ {x}

 is Scott open in X. To this end, let D be a directed subset of X with D ∈ Ux .    Then, D ∈ V and thus there is y ∈ D ∩ V . If y ≥ D, then y = D ∈ Ux . Otherwise, there is z ∈ D with z  y. Now, since D is directed, there is s ∈ D with y, z ≤ s. In fact, since z  y, y < s. Thus, s ∈ D ∩ (V − min(V )) ⊆ Ux as required. It follows that {Ux | x ∈ min(V )} is an open cover of V . By compactness it has a finite subcover. However, this implies that min(V ) is finite and thus V is a finitely generated up-set.  We note the following immediate corollary, which will be important in the next subsection when we characterize spectral domains. Corollary 7.37 Let X be a spectral dcpo. A subset of X is compact-open if, and only if, it is an finitely generated open up-set. In particular, a finite intersection of finitely generated open up-sets in X is a finitely generated open up-set. Proof The first statement holds by Lemma 7.36. The in particular statement follows because the compact-opens in a spectral space are closed under finite intersections.  Recall from Section 2.3 that a compact space is called coherent provided the intersection of any two compact-saturated subsets is again compact. Also recall from Exercise 6.3.5 that a topological space is a spectral space provided it is compact, T0 , coherent, sober, and has a base of compact-open subsets. Our goal in this subsection is to prove the following characterization of spectral dcpos, first announced in Erné (2009). Theorem 7.38 Let (X, τ) be a topological space. Then, X is a spectral dcpo if, and only if, the topology τ is coherent, sober, and has a base of finitely generated open up-sets. The proof of Theorem 7.38 will be a relatively straightforward combination of things we have already seen, with the notable exception of one crucial step, namely Proposition 7.41. We first prove a simple lemma on spaces having a base of finitely generated open up-sets. Lemma 7.39 Let X be a topological space. If X has a base of finitely generated open up-sets, then every compact-open in X is a finitely generated up-set. Proof Let U ⊆ X be compact-open. Since it is open, it is the union of all the open finitely generated up-sets it contains, and as U is compact, it is the union of finitely

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.2 Dcpos and Domains That Are Spectral

249

many such finitely generated open up-sets, ↑F1, . . . , ↑Fn . But generating up-sets is union preserving and thus U = ↑F1 ∪ · · · ∪ ↑Fn = ↑(F1 ∪ · · · ∪ Fn ).



We also require Rudin’s lemma, which is often useful in topology. Like the Alexander subbase theorem, it requires a non-constructive principle. Our statement and proof is based on Gierz et al. (1980, Lemma III-3.3). Lemma 7.40 (Rudin’s lemma) Let X be a poset and F a collection of finite non-empty subsets of X with the property that the collection {↑F | F ∈ F }

is filtering. Then, there is a directed set D ⊆ F so that D ∩ F  ∅ for each F ∈ F.

Proof Consider the collection S consisting of all subsets E ⊆ F with the following two properties: E ∩ F  ∅ for all F ∈ F ; and for all F, G ∈ F , if G ⊆ ↑F then E ∩ G ⊆ ↑(E ∩ F).

Notice that F ∈ S and thus S is non-empty. We show that any chain C in S has  a lower bound in S. To this end, let D := C. We claim that D has the required properties. First, we show that D ∈ S. Fix F ∈ F . Then, as C is a chain, so is {E ∩ F | E ∈ C}. Also, since F is finite, it is a finite chain, and by item (a), it consists of non-empty subsets of F. It follows that it has a minimum ∅  E0 ∩ F ⊆ F, where E0 ∈ C. It follows that D∩F = {E ∩ F | E ∈ C} = E0 ∩ F  ∅. (a) (b)

Now, consider F, G ∈ F with G ⊆ ↑F. As argued above, there are elements E1, E2 ∈ C so that D ∩ F = E1 ∩ F and D ∩ G = E2 ∩ G. Now, since C is a chain E1 and E2 are comparable, and we let E be the smaller of the two. Then, by minimality, we have D ∩ F = E1 ∩ F = E ∩ F

and

D ∩ G = E2 ∩ G = E ∩ G.

Finally, since E ∈ S, we have D ∩ G = E ∩ G ⊆ ↑(E ∩ F) = ↑(D ∩ F) as required. It follows, by Zorn’s lemma, that S contains minimal elements (i.e., maximal elements with respect to the order given by reverse inclusion). Before concluding, we make the following observation. Let E ∈ S and x ∈ X with the following property ∀F ∈ F

E ∩ F  ↑x.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

(7.3)

250

7 Domain Theory

We claim that E  = E ∩ (↑x) c is again in S. To this end, first note that (7.3) is equivalent to E ∩ F ∩ (↑x) c  ∅ for all F ∈ F , so that item (a) holds for E . For item (b), first notice that for any S ⊆ X we have ↑S ∩ (↑x) c ⊆ ↑[S ∩ (↑x) c ]. Now, if F, G ∈ F with G ⊆ ↑F then we have E  ∩ G = E ∩ G ∩ (↑x) c ⊆ ↑(E ∩ F) ∩ (↑x) c ⊆ ↑[E ∩ F ∩ (↑x) c ] = ↑(E  ∩ F). Now, let D ∈ S be minimal. Then removing elements from D results in a set not in S, thus the negation of (7.3) holds for each x ∈ D. That is, ∀x ∈ D ∃F ∈ F

D ∩ F ⊆ ↑x.

(7.4)

We show that this implies that D is directed. Let x, y ∈ D. Pick F, G ∈ F with D ∩ F ⊆ ↑x and D ∩ G ⊆ ↑y. Since {↑H | H ∈ F } is filtering, there is H ∈ F with H ⊆ ↑F ∩ ↑G. Finally, as D ∈ S, there is z ∈ D ∩ H. It follows that z ∈ D and z ∈ H ⊆ ↑F ∩ ↑G ⊆ ↑x ∩ ↑y and thus D is directed.



Proposition 7.41 Let X be a coherent sober space with a base of finitely generated open up-sets. Then, any Scott open set in the specialization order of X is open in X. Proof Let U ⊆ X be Scott open. We want to show that for each x ∈ U, there is a finitely generated open up-set ↑F with x ∈ ↑F ⊆ U. By contraposition, assume that x ∈ X is such that for any finite F ⊆ X, if ↑F is open and contains x, then F  U. We show that x  U. Consider the collection F := {F − U | F is finite, ↑F is open, and x ∈ ↑F}. We show that G = {↑G | G ∈ F } is filtering. For this purpose let F1, F2 be finite subsets of X with x ∈ ↑Fi and ↑Fi open for i = 1, 2. Since ↑F1 and ↑F2 are compactopen and X is coherent, we get that ↑F1 ∩ ↑F2 is compact-open. By Lemma 7.39, there exists a finite subset F ⊆ X such that ↑F1 ∩ ↑F2 = ↑F. It follows from this equality that x ∈ ↑F and ↑F is open. We show that F − U ⊆ ↑(Fi − U) for i = 1, 2. Since ↑F ⊆ ↑Fi we have F −U ⊆ ↑Fi . Let z ∈ F −U and let y ∈ Fi with y ≤ z. Since z  U and U is an up-set, we have y  U. So y ∈ Fi −U and thus F −U ⊆ ↑(Fi −U) for i = 1, 2. Therefore, by Rudin’s lemma, there is a directed subset  D⊆ F ⊆ X −U with D ∩ (F − U)  ∅ for each finite subset F of X with ↑F open and x ∈ ↑F. Now,  on the one hand, since U is Scott open, it follows that D  U, and on the other hand, since  D∈ {↑F | F is finite and x ∈ ↑F is open} = ↑x,

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.2 Dcpos and Domains That Are Spectral

it follows that x ≤



D and thus x  U as required.

251



With Proposition 7.41 in hand, we can now prove Theorem 7.38. Proof of Theorem 7.38 First, assuming X is a spectral dcpo, the compact-open sets form a base by definition of spectral spaces, but the compact Scott open sets are the finitely generated open up-sets by Lemma 7.36; thus, X has a base of finitely generated open up-sets. Now, assume that X is coherent, sober, and has a base of finitely generated open up-sets. The specialization order of a sober space is a dcpo by Proposition 7.6, and the topology is spectral because finitely generated up-sets are always compact in the Scott topology. Also, by Proposition 7.6, τ is contained in the Scott topology of  ≤τ . But the Scott topology of ≤τ is also contained in τ, by Proposition 7.41. Spectral Domains Having characterized the spectral dcpos, we now restrict our attention to domains whose associated space is spectral. Definition 7.42 Let X be a domain. We call X a spectral domain if the Scott topology on X is spectral. If X is a spectral domain, then X is a spectral dcpo, and X is compact, so by Corollary 7.37, X can only have a finite number of minimal elements. In any poset, it is clear that the union of two finitely generated up-sets is again a finitely generated up-set. On the other hand, it is not true in general that the finitely generated open up-sets of a domain are closed under binary intersection (see Exercise 7.2.2), while this must be true in a spectral dcpo, and thus in particular in spectral domains. For a general spectral dcpo, it is hard to understand in order-theoretic terms what it means for the finitely generated open up-sets to be closed under binary intersections, as we do not know which finite subsets F ⊆ X generate opens. However, if X is not just a dcpo but a domain, then the description of the compact-open subsets simplifies substantially, as we show now. Proposition 7.43 Let X be a spectral domain. Then, U ⊆ X is compact-open if, and only if, U = ↑F for some finite set F ⊆ K(X ) of compact elements of X. In particular, X is an algebraic domain. 

Proof For any x ∈ K(X ), we have ↑x = x, which is a Scott open set by Lemma

7.18. Thus, for any subset F ⊆ K(X ), ↑F = {↑x | x ∈ F} is open. Since any finitely generated up-set is compact, it follows that ↑F is compact-open when F ⊆ K(X ) is finite. For the converse, suppose that U ⊆ X is compact-open. Then, by Corollary 7.37, U is a finitely generated open up-set. Write F for the finite set of minimal elements

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

252

7 Domain Theory









of U. We show that F ⊆ K(X ). Let x ∈ F. By Lemma 7.18, since U is open, there is y ∈ U with y)x and thus, in particular, y ≤ x. Since x is minimal in U, it follows that y = x, so that x)x, that is, x is a compact element of X. For the “in particular” statement, let y ∈ X be arbitrary. We need to prove that   y, so it suffices y = ( y ∩ K(X )). Since X is a domain, we have that y = to prove that, for any y  )y, there exists a compact element x)y with y  ≤ x. Let y  )y be arbitrary. Since ) is interpolating by Lemma 7.18, pick z ∈ X with y  )z)y. Since z ∈ y , which is open, and since X is a spectral space, there is a compact-open set U such that z ∈ U ⊆ y . Since U is the upward closure of finitely many compact elements, there exists a compact element x ∈ K(X ) ∩U such that x ≤ z. Since x ∈ U, we have y  )x, so in particular y  ≤ x. Also, x ≤ z)y, so x)y, as required.  Recall that an element p in a lattice L is said to be join prime provided that, for  any finite F ⊆ L, p ≤ F implies that there exists a ∈ F with p ≤ a. Further, we say that L has enough join primes provided every element of L is the join of a finite set of join-prime elements. Also recall that, when X is a spectral space, “the distributive lattice dual to X” is by definition the lattice K O(X ) of compact-open sets of X, also see Figure 6.1. Corollary 7.44 Let X be a spectral domain. Then, the distributive lattice L dual to X is isomorphic to the poset of finitely generated up-sets of K(X ) with the settheoretic operations of intersection and union. In particular, L has enough join primes. Proof Consider the function from the poset of finitely generated up-sets of K(X ) to K O(X ) that sends T ⊆ K(X ) to the up-set ↑T of X . By Proposition 7.43, this is a well-defined surjective map, and it is clearly an order embedding. The poset of finitely generated up-sets of K(X ) clearly has enough join primes, as each ↑x with x ∈ K(X ) is join prime, and finitely generated up-sets of K(X ) are finite unions of such.  Corollary 7.44 shows that, for a spectral domain X, the poset K(X ) completely determines the lattice L dual to X, and thus also X itself. In particular, the poset K(X ) is isomorphic to J (L) op , as can be seen from Exercise 1.3.8 (also see Exercise 7.2.5). We now show that the spectral domains are in fact precisely the duals of the distributive lattices with enough join primes. Even though this theorem was first proved through the work of domain theorists in the 1980s and 1990s, culminating in the duality-theoretic formulation given here, a closely related result coming from a model- and set-theoretic tradition is Bekkali et al. (2007, Proposition 2.8).

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.2 Dcpos and Domains That Are Spectral

253

Theorem 7.45 Stone’s duality between spectral spaces and distributive lattices restricts to a duality between spectral domains and distributive lattices with enough join primes. Proof Corollary 7.44 establishes that the distributive lattice dual to a spectral domain has enough join primes. It remains to prove that the spaces of the form St(L), where L is a distributive lattice with enough join primes, are precisely the spectral domains. Let L be a distributive lattice with enough join primes and X its Stone dual. Then, X is clearly spectral. We show that it is a domain in its specialization order. Each join prime p of L gives rise to a principal prime filter x = ↑p and any such element x is a compact element of X. Now, let x, y ∈ X with x  y, then there is a ∈ L with a ∈ x but a  y. Since a is a finite join of join primes and x is a prime filter, there is p ∈ J (L) with p ≤ a and p ∈ x. Since a  y it follows that p  y. That is, ↑p ≤ x and ↑p  y and thus each element of X is a join of compact elements. It remains to show that this join is directed. Let p, q ∈ J (L) and x ∈ X with p, q ∈ x. Then, p ∧ q ∈ x and thus, as above for a, there is r ∈ J (L) with r ≤ p ∧ q and r ∈ x. It follows that ↑p, ↑q ≤ ↑r ≤ x in X and thus we have shown that X is a spectral domain.  The duality in Theorem 7.45 is centered around the poset P = K(X ) associated to the spectral domain X. An interesting feature of spectral domains is that there is a significant overlap between the space X and its dual lattice L, since this subposet P = K(X ) of X is isomorphic to the opposite of the poset J (L) of join-prime elements of L. It follows that this poset uniquely determines both X and L, since, in this case, X  Idl(P) and L  Dfin (Pop )  Ufin (P), the free join-semilattice on Pop , which, for this very special type of poset P, is not just closed under finite unions but also under finite intersections. We end this section by giving an order-theoretic characterization of the domains that are spectral in their Scott topology; that is, in Theorem 7.47, we will characterize the spectrality of a domain X entirely in terms of the poset P = K(X ). The characterizations of bifinite domains among spectral domains that we will give in Section 7.3 also make heavy use of this poset P, which lives on both sides of the duality. Definition 7.46 Let X be a poset, x ∈ X and M ⊆ X. Then, we will write M ≤ x provided x is a common upper bound of all the elements of M. That is, x belongs to the set UB(M) := {x ∈ X | ∀m ∈ M m ≤ x}. Further, we denote by MUB(M) the set of all minimal upper bounds of M. That is, MUB(M) := min(UB(M)). We are now ready to characterize, in order-theoretic terms, the spectral domains among the algebraic domains.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

254

7 Domain Theory

Theorem 7.47 Let X be an algebraic domain. Then, the following conditions on X are equivalent: (i) (ii)

the Scott topology on X is spectral; for all finite subsets F ⊆ K(X ), the following two properties hold: (1) (2)

(iii)

MUB(F) is finite and contained in K(X ) and UB(F) = ↑MUB(F); and

the finitely generated Scott-open up-sets of X are closed under finite intersections. 

Proof First, suppose X is a spectral domain and let F ⊆ K(X ) be finite. Then, ↑x = x is compact-open in X for each x ∈ F and thus the finite intersection ↑x UB(F) = x ∈F

is also compact-open. Thus, by Proposition 7.43, there is a finite antichain G ⊆ K(X ) such that UB(F) = ↑G. It follows that MUB(F) = min(UB(F)) = min(↑G) = G. That is, G = MUB(F) is a finite subset of K(X ) and UB(F) = ↑MUB(F), as required. Now, suppose that X is an algebraic domain satisfying the two properties in (ii). Recall from Exercise 7.1.13 that, in any algebraic domain, the finitely generated open up-sets are precisely the sets ↑G with G ⊆ K(X ) finite. Thus, by distributivity, it suffices to show that finite intersections of sets of the form ↑x, with x ∈ K(X ), are  finitely generated open up-sets. Let F ⊆ K(X ) be finite and U := x ∈F ↑x. Then, using (ii)(2), U = UB(F) = ↑MUB(F) and by (ii)(1), MUB(F) is a finite subset of K(X ), so that U is a finitely generated open up-set. Finally, to see that (iii) implies (i), recall that a topological space X is spectral if, and only if, it is sober and the collection of compact-open sets is a bounded sublattice of P (X ) which is a base for the topology, see Definition 6.2 and Exercises 6.1.3 and 6.3.5. Any algebraic domain is sober, and by Exercise 7.1.13, the finitely generated open up-sets form a base closed under finite unions, and any compact-open is a finitely generated open up-set. Now, since finitely generated open up-sets are closed under finite intersections by (iii), it follows that the finite intersection of compactopens is a finitely generated open up-set, and thus in particular compact again. Thus, the Scott topology on X is spectral. 

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.2 Dcpos and Domains That Are Spectral

255

Now, combining Proposition 7.43 and Theorem 7.47, we obtain the following order-theoretic characterization of spectral domains, which have also been called “2/3 bifinite” or “2/3 SFP” domains in the literature (see for example, Proposition 4.2.17 in Abramsky and Jung, 1994), because they satisfy two of the three properties that characterize bifinite domains, see Corollary 7.64 in Section 7.3. Corollary 7.48 Let X be a domain. Then, the Scott topology on X is spectral if, and only if, X is algebraic, and for all finite subsets F ⊆ K(X ) the following two properties hold: (a) (b)

MUB(F) is finite and contained in K(X ) and UB(F) = ↑MUB(F).

Remark 7.49 Posets P satisfying property (b) for every finite subset F ⊆ P are called MUB-complete in Abramsky and Jung (1994), and this property is sometimes also referred to as “property m.” Combining Corollary 7.48 with Exercise 7.1.11, it follows that spectral domains are, up to isomorphism, the posets of the form Idl(P), for P a MUB-complete poset in which MUB(F) is finite for any F ⊆ P; we will call such posets finitely MUB-complete. By Corollary 7.44 the dual bounded distributive lattice is Ufin (P) (see also Exercise 7.2.4). Definition 7.50 Let P be a poset. We say P is finitely MUB-complete if, for any finite subset F ⊆ P, MUB(F) is finite and UB(F) = ↑MUB(F). We note that a poset P is finitely MUB-complete if, and only if, the set Ufin (P) is a sublattice of P (P). To summarize the results in this section: A spectral domain X is always the domain of ideals of a finitely MUB-complete poset P, whose lattice of finitely generated up-sets is the dual distributive lattice of X.

Exercises for Section 7.2 Exercise 7.2.1 Let X be any T0 space and K ⊆ X a compact-saturated subset of X. Then, K = ↑ min(K ). Hint. Given x ∈ K, use Zorn’s lemma on the collection of closed sets ↓y, as y ranges over elements of ↓x ∩ K, equipped with the reverse inclusion order, to show that every element of K is above a minimal one. Exercise 7.2.2 Let X be a poset and M, N ⊆ X. (a) (b) (c)

Show that ↑M ∪ ↑N = ↑(M ∪ N ) and conclude that the finitely generated open up-sets of any topological space are closed under binary unions.

Show that ↑M ∩ ↑N = {UB({m, n}) | m ∈ M and n ∈ N }. Give an example of a topological space, and finitely generated open up-sets U and V so that U ∩ V is not a finitely generated up-set.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

256

(d) (e)

(f)

7 Domain Theory

Give a domain whose associated space gives an example as in (c). Give examples showing that, even if UB(M) is non-empty, MUB(M) may be empty, and even if MUB(M) is non-empty, we may not have UB(M) = ↑MUB(M). Show that if X = Idl(P) and F ⊆ P is finite, then MUB(F) ⊆ P.

Exercise 7.2.3 Prove that a poset P is finitely MUB-complete if, and only if, any finite intersection of finitely generated up-sets of P is finitely generated. Conclude that this happens if, and only if, Ufin (P) is a bounded sublattice of P (P). Exercise 7.2.4 Let X be a spectral domain. Show that the following three are isomorphic distributive lattices. (a) (b) (c)

K O(X ), the distributive lattice dual to X. Ufin (K(X )), the lattice of finitely generated up-sets of the poset K(X ), with the inclusion order. The poset reflection of the set Pfin (K(X )) of finite subsets of K(X ), equipped with the preorder given by F  G ⇐⇒ ∀y ∈ G ∃x ∈ F (x ≤ y). (The poset reflection was defined in Exercise 1.1.5.)

Exercise 7.2.5 Let X be a spectral domain and let L be its lattice of compact open subsets. Prove that J (L) op is isomorphic to K(X ). Hint. You can describe an isomorphism directly: Send p ∈ K(X ) to the join-prime element ↑p of L.

7.3 Bifinite Domains Bifinite domains 2 are mathematical structures obtained as limits of finite “embedding projection pairs,” which we will introduce below. Within domains, one can show that these are actually bilimits, that is, simultaneously limits of the projections and colimits of the embeddings in the finite embedding projection pairs. Limits of finite posets clearly make profinite posets, that is, Priestley spaces (recall Example 5.35), pertinent to the subject, but it is only over time that the duality-theoretic point of view came into focus, fully expressed in Abramsky (1991). Here we introduce bifinite domains as a subclass of spectral spaces, and we consider them from 2

SFP domains were introduced in Plotkin (1976) as certain profinite posets, along with the powerdomain construction, under which they are closed. The name “SFP” is an acronym for “sequences of finite posets.” SFP domains are further required to be ω-algebraic, that is, the set of compact elements needs to be countable, and pointed, that is, it needs to have a least element. Bifinite domains may be viewed as the generalization of SFP domains where the ω-algebraicity and pointedness are dropped. The minimal upper bounds point of view that we discuss here was developed by Smyth (1983b) and in the thesis of Gunter (1985). Smyth showed that SFP is the largest Cartesian closed category of pointed ω-algebraic domains.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.3 Bifinite Domains

257

a duality-theoretic point of view. In Corollary 7.64 we give a characterization of bifinite domains via MUB-completeness and in Theorem 7.66 we give a dual characterization in terms of a condition of “conjunctive closure” on distributive lattices, bringing us full circle, showing that these are precisely the structures studied in Abramsky (1991). A central notion in the definition of bifiniteness is that of a (finite) embedding projection pair, which we introduce now. This definition will be applied both to distributive lattices and to spectral domains. In the following definition, the category C can be thought of as either DL or Spec, or in fact any category equipped with a faithful functor to Poset. Definition 7.51 Let C be a concrete category in which each object is equipped with a partial order and each morphism is order preserving. We say that a pair of morphisms C

g

D of C is an embedding projection pair (EPP) provided

f

( f , g) is an adjoint pair, f is injective, and g is surjective. Here, f is called the embedding and g is called the projection of the pair. Further, such an EPP is said to be finite if C is finite, and we call it an EPP “of” D. Recall from Exercise 1.1.9 that, for an adjoint pair between posets, the left adjoint is injective if, and only if, the right adjoint is surjective; so, for an adjunction to be an EPP, it suffices to check one of the two conditions (also see Exercise 7.3.1). We start by characterizing the finite EPPs for distributive lattices and spectral spaces in terms of substructures, see Propositions 7.54 and 7.55. Notation In what follows, we often consider pairs of functions between spectral spaces and distributive lattices. If X and Y are spectral spaces, then we use the notation X

p e

Y for a pair of spectral maps between these spaces. We then

call the dual pair of homomorphisms the pair of functions K

h i

L , where K

is the lattice dual to X, L is the lattice dual to Y , i is the homomorphism dual to p, and h is the homomorphism dual to e. p

Lemma 7.52 Let X X and Y and let K

e h i

Y be a pair of spectral maps between spectral spaces L be the dual pair of homomorphisms. Then, e is lower

adjoint to p if, and only if, i is lower adjoint to h. Proof The function e is lower adjoint to p if, and only if, e◦p ≤ idY and idX ≤ p◦e (see Exercise 1.1.8). The dual of the composite map e ◦ p is i ◦ h and the dual of idY is id L , so, using Remark 7.3, e ◦ p ≤ idY is equivalent to i ◦ h ≤ id L . Similarly, idX ≤ p ◦ e is equivalent to idK ≤ h ◦ i. Thus, e is lower adjoint to p if, and only

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

258

7 Domain Theory

if, i ◦ h ≤ id L and idK ≤ h ◦ i. This, in turn, is equivalent to saying that i is lower adjoint to h.  We have the following corollary of Lemma 7.52 and Exercise 7.3.1. Corollary 7.53 The Stone dual of an embedding projection pair on either side of the duality is an embedding projection pair on the other side, and the dual of a finite EPP is finite. Note that the dual of the embedding on either side is the projection on the other. The following propositions identify the nature of finite EPPs on either side of Stone–Priestley duality. Recall that a map f between topological spaces is said to be an embedding provided it is injective and the inverse function f −1 : im( f ) → X is also continuous. If Y is a spectral space and X ⊆ Y is a subspace of Y which is itself a spectral space in the subspace topology, then we call the inclusion map e : X → Y a spectral subspace embedding. The following proposition allows us to see EPPs on a spectral space Y as certain spectral subspaces X of Y . Recall that we view a spectral space as a poset in its specialization order. Proposition 7.54 Let Y be a spectral space and e : X → Y a spectral subspace embedding. Then, the following two conditions are equivalent: (i) (ii)

the function e has an upper adjoint p, and p is spectral, and the subspace X has the following two properties: (1) for all y ∈ Y , the down-set ↓y ∩ X is principal; and (2) for all U ⊆ X compact-open, the up-set ↑U is open in Y . If, in addition, X is finite, then it suffices that ↑x is open in Y for all x ∈ X.

Proof Note first that, essentially by definition, e has an upper adjoint if, and only if, for every y ∈ Y , ↓y ∩ X has a maximum. Thus, (ii)(1) is equivalent to the existence of an upper adjoint. Now, suppose the upper adjoint exists, and consider the map p : Y → X, y → max(↓y ∩ X ). We show that (ii)(2) is equivalent to p being a spectral map. Notice that, for any up-set U of X we have p−1 (U) = ↑U. Thus, p being spectral is the statement that ↑U is compact-open whenever U ⊆ X is compact-open. Also note that U compact in X implies U compact in Y , which, in turn, implies that ↑U is compact in Y . Thus, we just need to know that ↑U is open in Y . If, in addition, X is finite, then every compact-open of X is a finite union of principal up-sets ↑X x and thus it suffices to consider these.  Proposition 7.55 Let L be a bounded distributive lattice and i : K → L a bounded sublattice inclusion. Then, the following conditions are equivalent: (i)

the homomorphism i has an upper adjoint that is also a homomorphism; and

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.3 Bifinite Domains

(ii)

259

the sublattice K has the following two properties: (1) (2)

for all b ∈ L, ↓b ∩ K is a principal down-set; and for any prime filter F of K the filter ↑F in L is again prime.

Proof Again, the existence of an upper adjoint is equivalent to (ii)(1) since the upper adjoint of the embedding must be given by h(b) = max(↓b∩K ). We show that (ii)(2) is equivalent to this upper adjoint being a bounded lattice homomorphism. Consider the map h : L → K, b → max(↓b ∩ K ). Since h is the upper adjoint of the inclusion, it preserves all existing meets. Also, clearly h(⊥) = ⊥. We show that (ii)(2) is equivalent to h preserving binary joins. For this purpose, assume h preserves binary joins and let F be a prime filter of K. Then, as F is proper, so is ↑F. If b1 ∨ b2 ∈ ↑F in L, then there is a ∈ F with a ≤ h(b1 ∨ b2 ) = h(b1 ) ∨ h(b2 ) in K. Now, since F is a prime filter, it follows that h(b1 ) ∈ F or h(b2 ) ∈ F and, since h(bi ) ≤ bi , it follows that b1 ∈ ↑F or b2 ∈ ↑F, as required. Conversely, suppose that (ii)(2) holds. We want to show that h preserves binary joins. Since h is order preserving we have h(b1 ) ∨ h(b2 ) ≤ h(b1 ∨ b2 ). Since any filter is the intersection of the prime filters containing it (see Exercise 3.1.17), it suffices to show that any prime filter of K containing h(b1 ∨ b2 ) must contain h(b1 ) ∨ h(b2 ). To this end, let F be a prime filter of K with h(b1 ∨ b2 ) ∈ F. Since h(b1 ∨ b2 ) ≤ b1 ∨ b2 , we have b1 ∨ b2 ∈ ↑F and by (ii)(2), it follows that bi ∈ ↑F for i = 1 or 2. Now, bi ∈ ↑F implies there is a ∈ F with a ≤ bi , and as a ∈ K, it follows that a ≤ h(bi ) and thus  h(bi ) ∈ F. That is, h(b1 ) ∨ h(b2 ) ∈ F as required. Just as for spectral spaces, Proposition 7.55 allows us to see EPPs on a distributive lattice L as certain sublattices K of L, namely those satisfying the two conditions of Proposition 7.55(ii). Interestingly, Corollary 7.53 then yields a duality between certain subspaces and certain sublattices, rather than the usual matching in duality of subs and quotients. This is of course because these subs are also quotients, but viewing this as a duality between sublattices and subspaces is interesting relative to a phenomenon we will meet in our second application in Chapter 8, where certain residuated Boolean algebras are dual to certain profinite algebras, namely profinite monoids. We will be particularly interested in finite EPPs. Since any inclusion of a finite (join-semi-)lattice in a lattice has an upper adjoint, and since any prime filter in a finite lattice is principal, generated by a join-prime element, we obtain from Proposition 7.55 the following much simpler description of finite EPPs in a distributive lattice. Corollary 7.56 Let L be a distributive lattice and K ⊆ L a finite bounded sublattice, with i : K → L the inclusion map. Then, the following conditions are equivalent:

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

260

(i) (ii)

7 Domain Theory

there exists a bounded lattice homomorphism h : L → K such that (i, h) is an EPP in the category DL; and J (K ) ⊆ J (L).

Definition 7.57 Let L be a distributive lattice. We say that a finite bounded sublattice K of L is a finite EPP-sublattice if it satisfies the equivalent conditions of Corollary 7.56. Note that, if K is a finite EPP-sublattice of L, then K has “enough join primes,” in the sense that every element of K is a finite join of join-prime elements (see Lemma 1.20), and every join prime of K is also a join prime of L. However, an infinite lattice L need not have any join-prime elements at all (see Example 3.1), and such a lattice does not have any finite EPP. Thus having “enough” finite EPPs is special. For L to be bifinite, in addition, we require that the finite EPP-sublattices of L form a directed diagram. Here, recall that “directed” means that for any two finite EPP-sublattices K1 and K2 of L, there is a finite EPP-sublattice K0 of L which contains both K1 and K2 . Definition 7.58 Let X be a spectral space, L its dual lattice. We say that X and L are bifinite provided either and then both of the following two equivalent conditions are satisfied: (i) (ii)

X is the projective limit in Spec of the projections of its finite EPPs; and L is the directed colimit in DL of the embeddings of its finite EPPs.

In light of Definition 7.57 and the fact that directed colimit of a collection of sublattices of a distributive lattice can be computed as a union (see Example 5.34), an equivalent definition for a distributive lattice L to be bifinite is that the finite EPP-sublattices of L form a directed diagram whose union is equal to L. Remark 7.59 Definition 7.58 is not identical to any of the standard ones. The definition is usually given for algebraic domains, and not for spectral spaces, but as we will observe shortly, any bifinite spectral space is an algebraic domain. That is, the spaces we call bifinite are exactly the same objects as the bifinite domains of the domain theory literature. Remark 7.60 Given a bifinite lattice or spectral space, we have a directed diagram of EPPs for that structure, and thus we can form two different diagrams: the diagram consisting of the embedding parts of the diagram of finite EPPs and the diagram consisting of the projection parts of the diagram of finite EPPs for the given structure. By definition, the colimit of the diagram of embeddings in a bifinite lattice yields the bifinite lattice itself. Dually the limit of the diagram of projections of the finite EPPs in a bifinite spectral space yields the bifinite space. On the other hand, if we

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.3 Bifinite Domains

261

take the limit in DL of the diagram of projections in a bifinite lattice, in general, we will get a bigger lattice, which has the colimit as a sublattice, and similarly, if we take the colimit in Spec of the diagram of embeddings in a bifinite spectral space we will get a bigger space, which has the limit as a quotient (see Exercise 7.3.3). However, if we embed DL in Frame via the ideal completion, as in Figure 6.1, and take the full subcategory of Frame it generates, then we get a category equivalent to the category DL j−approx of distributive lattices with join-approximable relations (see Exercise 6.3.7) and in this category we have a morphism from the limit in DL of the diagram of projections in a bifinite lattice to the bifinite lattice L itself. This makes it possible to show that we have coincidence of the colimit of the injections and limit of the projections, thus showing that in this category the limit object L is so-called bifinite, that is, simultaneously the colimit and the limit of its diagram of finite EPPs (for more details see Exercise 7.3.3). Finally, by duality, the same is true if we consider bifinite spectral spaces with arbitrary continuous functions rather than just spectral maps, see also Abramsky and Jung (1994, Section 3.3.2). Corollary 7.61 Let X be a spectral space and L its dual lattice. If X and L are bifinite then X is a spectral domain, or, equivalently, L has enough join primes. Proof If a distributive lattice L is bifinite, then every element a ∈ L is contained in a finite EPP-sublattice K. In K, a is a finite join of join-prime elements, and, by Corollary 7.56, these join-prime elements are also join prime in L. Thus, every element of L is a finite join of join-prime elements. Now, by Theorem 7.45 this is equivalent to X being a spectral domain.  We will now give an order-theoretic characterization of bifinite spectral spaces. By Corollary 7.61, if X is bifinite with dual lattice L, then X is a spectral domain. That is, by Remark 7.49, the poset P = K(X ), which is isomorphic to J (L) op , is finitely MUB-complete. We further have X  Idl(P) and L  Ufin (P) (see Exercise 7.2.4). In Proposition 7.63, we will characterize bifinite domains in terms of order-theoretic properties of P. Definition 7.62 Let X be a poset and F ⊆ X. We say that F is MUB-closed provided, for all G ⊆ F, we have MUB(G) ⊆ F. The MUB-closure of F ⊆ X is the inclusion-minimal F  ⊆ X such that F ⊆ F  and F  is MUB-closed. Proposition 7.63 Let P be a finitely MUB-complete poset and let X := Idl(P) be the corresponding spectral domain. Then, a finite subspace F ⊆ X gives rise to an EPP in the category Spec if, and only if, F ⊆ P and F is MUB-closed. Furthermore, the spectral domain X is bifinite if, and only if, the MUB-closure of any finite subset of P is finite.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

262

7 Domain Theory

Proof A finite subspace F ⊆ X gives rise to an EPP if, and only if, it satisfies the equivalent conditions of Proposition 7.54. The finite case in condition (ii)(2) of that proposition is equivalent to F ⊆ P. We show that (ii)(1) is equivalent to F being MUB-closed. Let F ⊆ P be finite and suppose that max(↓x ∩ F) exists for each x ∈ X. Let G ⊆ F and let p ∈ MUB(G) and consider the set ↓p ∩ F. By our assumption, there is p  = max(↓p ∩ F). Also, clearly G ⊆ ↓p ∩ F, so p ∈ UB(G). Since P is MUB-complete, there is p ∈ MUB(G) with p ≤ p. Thus, we have p ≤ p ≤ p with both p, p ∈ MUB(G). It follows that p = p = p and thus p ∈ F since p ∈ F. That is, F is MUB-closed. For the converse, again let F ⊆ P be finite, suppose F is MUB-closed, and let x ∈ X. Further, set G := ↓x ∩ F. Since x ∈ UB(G), there is a p ∈ MUB(G) ⊆ F with p ≤ x. But then p ∈ G and thus p = max(G) = max(↓x ∩ F), as required. Finally, X is bifinite if, and only if, every finite subset F of P is contained in a finite subset F  which gives rise to an EPP, which we showed happens if, and only if, F  ⊆ P and F  is MUB-closed. Thus, X is bifinite if, and only if, the MUB-closure of any finite subset of P is finite.  Corollary 7.64 Let X be an algebraic domain. Write P := K(X ), so that X  Idl(P). Then, X is bifinite if, and only if, for all finite subsets F of P, the following three properties hold: (a) MUB(F) is finite, (b) UB(F) = ↑MUB(F), and (c) the MUB-closure of F is finite. A poset P is called a Plotkin order if it satisfies the three properties in Corollary 7.64 for all finite subsets F of P. Thus, Corollary 7.64 shows that an algebraic domain X is bifinite if, and only if, the poset P = K(X ) is a Plotkin order. Note that the properties (a) and (b) in Corollary 7.64 are equivalent to saying that the corresponding domain X = Idl(P) is spectral, see Corollary 7.48. We finish the subsection with a characterization of the lattices dual to bifinite domains. For this purpose, we define the notion of a subset of a lattice being conjunctively closed. Definition 7.65 Let L be a lattice and S ⊆ L. We say that S is conjunctively closed provided, for each finite subset F ⊆ S, there is a finite G ⊆ S such that   F= G. Theorem 7.66 Stone’s duality between Spec and DL restricts to a duality between bifinite domains and bounded distributive lattices L with enough join primes satisfying the following property: for every finite S ⊆ J (L), there exists S  ⊆ J (L) finite, such that S  ⊇ S and S  is conjunctively closed.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.3 Bifinite Domains

263

Proof By Theorem 7.45 spectral domains are dual to bounded distributive lattices L with enough join primes, so we just need to verify that the additional property in the theorem corresponds dually to condition (c) in the definition of bifinite domains. This is left as Exercise 7.3.2.  Example 7.67 We show that (N, |), the divisibility lattice of N, is a bifinite lattice. Recall that we computed the Priestley dual space of (N, |) in Example 3.18. To prove that the lattice L := (N, |) is bifinite, let n ∈ N be arbitrary. Denote by Dn the subset of natural numbers that are 0 or divisors of n, that is, the down-set of n in L, with the top element 0 added. This is a finite bounded sublattice of L; we apply Corollary 7.56 to show that it is an EPP-sublattice. Let j be join prime in Dn . Then, j |n and we show that j must in fact be a positive power of a prime number: since 1 is the bottom element of Dn , j  1, and we may write j as the join of p1k1 , . . . , prkr , for some r ≥ 1, prime numbers p1, . . . , pr and k1, . . . , kr ≥ 1. Then, each piki divides j, and hence, since j ∈ Dn , piki is also an element of Dn . Moreover, by the assumption that j is join prime and the fact that each piki is incomparable to k

p j j , we must have r = 1. Thus, j is a power of a prime number, and therefore also join prime in L. Recall that we gave an example of a quotient space of the dual of (N, |) in Example 3.30. For any n ∈ N, the finite EPP-sublattice Dn defined here dually yields a different finite quotient of the Priestley dual space of (N, |). Exercises for Section 7.3 Exercise 7.3.1

Let X

spaces X and Y and let K

p

Y be a pair of spectral maps between spectral

e h i

L be the dual pair of homomorphisms. Show

that all of the following are equivalent: (i) (ii) (iii) (iv)

e is injective; p is surjective; i is injective; and h is surjective.

Further show that, if these hold, then e and i are embeddings. Exercise 7.3.2 Let X be a spectral domain, P := K(X ), and L the lattice dual to X. Show that the following are equivalent: (i) (ii)

for every finite F ⊆ P, the MUB-closure of F is finite; and for every finite S ⊆ J (L), there exists finite S  ⊆ J (L) such that S  ⊇ S, and S  is conjunctively closed.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

264

7 Domain Theory

Hint. Use the isomorphism between J (L) op and P and the fact that, for a finite collection F ⊆ P, if G = MUB(F), then  ↑p = ↑q. UB(F) = p ∈F

q ∈G

Exercise 7.3.3 Let C be a concrete category in which each object is equipped with a partial order and each morphism is order preserving. Let   g C D f

be a directed diagram of EPPs in C. Further, let C f be the colimit in C of the embeddings of the diagram and Cg the limit in C of the projections of the diagram. (a) (b)

Show that there is a unique morphism C f → Cg in C. Consider now the particular case C = DL. Let L be a bifinite lattice and let   h L K i be the associated diagram of finite EPPs. Prove that the colimit L i of the embedding parts of the diagram is isomorphic to L, and that the homomorphism L → L h , obtained from the previous item by composing with this isomorphism is an embedding. Conclude also that, by Stone duality, if X is a bifinite spectral space and   p X Y e

(c)

(d)

(e)

is the associated diagram of finite EPPs, then the limit X p of the projection parts is homeomorphic to X, and the spectral map Xe → X, obtained from the previous item by composing with this homeomorphism, is a quotient map. Prove that, if L is a bifinite lattice and X is the dual spectral space, then with the notations of the previous item, L h is dual to Xe and the morphism L → L h is dual to the spectral map Xe → X. Now, consider the situation where L = Ufin (P), where P is a Plotkin order. Show that L h  U (P) and conclude that L and L h are not necessarily isomorphic. Still under the assumption that L = Ufin (P), where P is a Plotkin order, show that the relation R = {(U, V ) ∈ U (P) × Ufin (P) | U ⊇ V } is a join-approximable relation from the lattice L h  U (P) to the lattice L = Ufin (P), and conclude that L = Ufin (P) is the bilimit of the diagram

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.4 Domain Theory in Logical Form

265

of EPPs for L in the category of distributive lattices with join-approximable relations. Exercise 7.3.4 This exercise is about the lattice N with the divisibility order. We already studied this lattice as a running example for Priestley duality in Chapter 3, see Examples 3.6, 3.18, and 3.30, also see Example 7.67 in this section. Here, we consider the spectral space X dual to (N, |), and its specialization order. Recall from Remark 7.3 that the specialization order on X is the opposite of the order on the Priestley space dual to (N, |), depicted in Figure 3.4. (a) (b) (c) (d)

Show that for any x ∈ X, the principal up-set ↑x is either finite or equal to X. Show that the Scott topology on the spectral space X coincides with the Alexandroff topology. Deduce that X is an algebraic domain in which all elements are compact, and that the poset K(X ) = X is finitely MUB-complete. Identify those finite subsets of X which correspond to finite EPPs, and show that X is bifinite. 7.4 Domain Theory in Logical Form

In this section, we analyze the DTLF program in the setting of bifinite domains as in Abramsky (1991), emphasizing the duality-theoretic aspects. We will illustrate DTLF by considering duality for the function space construction and a domain equation based on this constructor. Bifinite Domains Are Closed under Function Space In this subsection, we use the duality developed for function spaces in Section 6.4 to show that the category Bif of bifinite domains with Scott continuous functions is closed under the function space construction. Recall that, for topological spaces X, Y , the set [X, Y ] of continuous functions from X to Y is equipped with the compact-open topology: A subbase is given by the sets of functions K → U := { f ∈ [X, Y ] | f [K] ⊆ U}, for K ⊆ X compact and U ⊆ Y open, see Definition 6.26. We aim to prove here (Theorem 7.69) that this space [X, Y ] is bifinite if both X and Y are. For this purpose, we will make crucial use of the fact that this space is given by the smallest congruence that is join preserving at primes, the property that we identified in Definition 6.32. Recall that, if X and Y are bifinite, then in particular they are spectral domains by Corollary 7.61 and thus, by Corollary 6.34, [X, Y ] is a spectral space. We will need to prove that [X, Y ] can still be approximated by finite EPPs, if both X and Y have this property. To this end, we first prove a lemma about lifting finite EPPs to function spaces.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

266

7 Domain Theory

Lemma 7.68 Let X, Y be two spectral spaces such that [X, Y ] is a spectral space and let X , Y  be finite posets, equipped with the Alexandrov topology. If X

pX eX

X and Y 

pY eY

Y are EPPs, then the functions e : [X , Y ] → [X, Y ] : p f

→ eY f  pX

pY f eX ←

f

are spectral, and form an EPP between [X , Y ] and [X, Y ]. Moreover, for any  −1  up-sets K  ⊆ X , U  ⊆ Y , p−1 (K  → U  ) = p−1 X (K ) → pY (U ). Proof Let X 

pX eX

X and Y 

pY eY

Y be finite EPPs. We define

e : [X , Y ] → [X, Y ] : p f

→ eY f  pX

pY f eX ←

f.

Note that, for any f  ∈ [X , Y ], we have pe( f  ) = pY eY f  pX eX = idY  f idX  = f , showing that p ◦ e = id[X ,Y ] . Also, for any f ∈ [X, Y ], we have ep( f ) = eY pY f eX pX ≤ idY f idX = f , showing that e ◦ p ≤ id[X,Y] . Thus, (e, p) is an adjoint pair of order-preserving functions between the underlying posets, e is injective, and p is surjective. It remains to show that e and p are spectral maps. In order to prove that e is continuous, and hence spectral because [X , Y ] is finite, it suffices to show that e−1 (K → U) is an up-set for every K ⊆ X compact and U ⊆ Y open. Recall that the specialization order on a function space coincides with the pointwise order when the domain is locally compact (see Exercise 6.4.3(c)), and so certainly in this case. Let K ⊆ X be compact and U ⊆ Y be open. Suppose that e( f  ) ∈ K → U for some f  ∈ [X , Y ], and that f  ≤ g for some g ∈ [X , Y ]. We show that e(g) = eY gpX ∈ K → U. Let x ∈ K be arbitrary. Note that f  pX [K] ⊆ eY−1 (U) by assumption, and that f  pX (x) ≤ gpX (x). Thus, since eY−1 (U) is an up-set by continuity of eY , it follows that gpX (x) ∈ eY−1 (U), that is, eY gpX (x) ∈ U, as required. Note that the fact that p is spectral will follow from the “moreover” statement, because both pX and pY are spectral, and the sets of the form K  → U , with K  and U  up-sets of X  and Y , respectively, form a base for the topology on [X , Y ]. To prove the moreover statement, let K  ⊆ X  and U  ⊆ Y  be upsets and let f ∈ [X, Y ] be arbitrary. We will show that p( f ) ∈ K  → U  if,

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.4 Domain Theory in Logical Form

267

 −1  and only if, f ∈ p−1 X (K ) → pY (U ). For the left-to-right direction, suppose that p( f ) ∈ K  → U  and let x ∈ X be such that pX (x) ∈ K . Then, using the adjunction between eX and pX , and the fact that f is order preserving,

pY f (x) ≥ pY f eX pX (x) = p( f )(pX (x)), and the latter is an element of U  since p( f ) ∈ K  → U  by assumption, so that  pY f (x) ∈ U  since U  is an up-set. Conversely, suppose that f ∈ p−1 X (K ) → pY−1 (U  ) and let x  ∈ K  be arbitrary. Using the fact that pX eX (x  ) = x , so that   −1    eX (x  ) ∈ p−1 X (K ), we get f e X (x ) ∈ pY (U ), which shows that p( f )(x ) ∈ U , as required.  Theorem 7.69 Let X and Y be bifinite domains. Then, [X, Y ] is a bifinite domain. Proof Denote by L and M the bifinite distributive lattices of compact-opens of X and Y , respectively. Note that L in particular has enough join primes, by Corollary 7.61. Thus, Corollary 6.34 applies, and we obtain the fact that the space [X, Y ] is spectral, and its distributive lattice of compact-opens is (up to isomorphism) the lattice K := F→ (L, M)/ϑjpp . By duality, it is thus equivalent to show that this lattice K is bifinite, that is, that every element of K lies in some finite EPP-sublattice. Since the elements of the form a → b, with a ∈ L and b ∈ M, generate the lattice K, it suffices to show the result for these elements. Let a ∈ L and b ∈ M. Since L and M are bifinite, pick finite EPP-sublattices L  and M  of L and M, respectively, with dual EPPs X 

pX eX

X and Y 

pY eY

Y . Combining Lemma 7.68 with

Lemma 7.52, the lattice of compact-opens of [X , Y ] is isomorphic to a finite EPPsublattice K  of K, with embedding given by p−1 and projection given by e−1 . This sublattice K  in particular contains the element a → b because p−1 (a → b) is −1 equal to p−1 X (a) → pY (b). Thus, K is a bifinite lattice. Finally, since X and Y are bifinite, they are spectral domains and thus carry their Scott topologies. It follows in particular that [X, Y ] is the set of Scott continuous functions from X to Y . Finally, since [X, Y ] is bifinite, it is also a spectral domain by Corollary 7.61, so that the spectral topology on [X, Y ] coincides with the Scott topology of its specialization order.  Variations on a Domain Equation In the above proof of Theorem 7.69, we proved a fact about the type constructor of function spaces of domains by transferring it to a dual construction on the corresponding distributive lattices. This is the fundamental idea of DTLF (Abramsky, 1987, 1991): All of the domain or type constructors of interest (sums, products, various forms of power domains, functions spaces, etc.) are dual to endofunctors on the category of bifinite lattices that automatically preserve directed colimits

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

268

7 Domain Theory

of sequences since they are given “algebraically,” that is, freely by generators and relations. It follows that any composition F of these endofunctors also preserves directed colimits. This allows one in particular to build fixed points, that is, solutions of equations of the form X  F (X ), by iterating F, starting from a morphism A → F ( A) (i.e., a co-algebra for the functor F). A canonical such fixed point is the least fixed point of F, which can be constructed as the colimit of the following sequence, starting from the free distributive lattice on the empty set, 2: e0

e1

e2

2 −→ F (2) −→ F 2 (2) −→ · · · .

(7.5)

Here, F is any directed-colimit-preserving functor on DL, e0 is the unique morphism from 2 to F (2), given by the fact that 2 is initial in DL, and en+1 := F (en ). If, in addition, the maps of the sequence consist of EPPs then we can conclude that the fixed point obtained as the colimit of (7.5) is again bifinite. We now illustrate these ideas by considering the most classical example of a domain equation, namely X  [X, X]. One may first look for the least solution of this equation, but, as classical as this case is, it is also an anomaly, in that the least solution is not so interesting, since it is the one-element space. On the dual side, this is reflected by the fact that the iterative sequence (7.5), applied in the case where F is the functor sending a lattice L to F→ (L, L)/ϑjpp , never “gets off the ground,” that is, the very first map e0 : 2 → F→ (2, 2)/ϑjpp is already an isomorphism. This follows from the observation that [1, 1]  1 and Corollary 6.34, but it is also not hard to give an elementary algebraic proof directly from the definition of the lattice F→ (2, 2)/ϑjpp : notice first that already in F→ (2, 2), we have the equalities 0 → 0 = 1,

0 → 1 = 1 , and

1→1=1

because an implication-type operator by definition sends any pair with ⊥ in the first coordinate or  in the second coordinate to ; indeed, these are the “empty set” instances of the schemes (6.2) and (6.3) that define F→ (L, M). Moreover, since ϑjpp makes → preserve joins at primes, and 1 is a join-prime element in 2, we must have 1 → 0 = 0 in F→ (2, 2)/ϑjpp . Thus, all four generators of F→ (2, 2)/ϑjpp are equal to 0 or 1. For the rest of this section, let FEnd denote the functor on bifinite lattices that sends L to F→ (L, L)/ϑjpp , which is the dual lattice of the endomorphism type constructor X → [X, X] on bifinite domains, by Corollary 6.34. The previous paragraph shows that the least fixpoint of FEnd is 2, so we can not directly use the construction in (7.5). There are, essentially, two ways around this problem. One is to change the functor with which we induct; the other is to start the induction “higher up.” Among the amended functors, a minimal and very natural choice is to consider, for X a bifinite domain, the domain T (X ) = ⊥ ⊕ [X, X], usually denoted [X, X]⊥ , which takes the

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.4 Domain Theory in Logical Form

269

function space and adds a new bottom. Clearly, the dual is L → FEnd (L) ⊕ 1, which adds a new top. This functor is natural from the programming languages point of view and has been studied extensively, because X  [X, X]⊥ is the domain equation corresponding to what is known as the “lazy lambda calculus,” in which lambda terms are identified with their so-called weak head normal form (Abramsky, 1990; Abramsky and Ong, 1993). In this case, the canonical solution, that is, the least fixed point of the sequence (7.5), is non-trivial but its function space is a retract of X, and not isomorphic to X. This construction is the subject of Example 7.71. From the theoretical point of view, it is of course also important to show that there exist non-trivial, on-the-nose, solutions to the equation X  [X, X]. The way to obtain this is to stick with the original functor FEnd as the functor we iterate, but to start the iterative process in (7.5) with a bigger lattice than 2. This is the approach taken by Scott at the very start of domain theory, see, for example, Scott (1972, 1980). This however inherently involves an ad hoc choice of a starting point for the iterative sequence. In Smyth and Plotkin (1982), this approach is cast in a general scheme based on solving the domain equation in a comma category. The solution of X  [X, X] starting from Sierpinski space, or, on the lattice side, from the three-element chain, is the subject of Example 7.70. It is also worth mentioning the treatment in the book of Lambek and Scott (1986) on categorical logic, which solves a three-way equation X  [X, X]  X × X. More generally, modeling the lambda calculus and associated programming languages is a vast field in theoretical computer science which goes far beyond the confines of duality theory, see, for example, the books Barendregt (2014) and Barendregt and Manzonetto (2022), and also see Manzonetto and Salibra (2008) for a survey on universal algebra in lambda theories. Example 7.70 We want a solution to the domain equation X  [X, X], starting the iterative process from the Sierpinski space S. In dual form, this means we want a solution to the equation L  FEnd (L) within bifinite lattices, starting from the threeelement chain, which we will denote by 3. As we have explained above, in order to get started, we need an embedding 3 → FEnd (3) or, dually speaking, a projection [S, S] → S. Also, as we shall see, it will be important that these morphisms are part of embedding-projection pairs. So let’s first see whether this is possible. As the calculation is easier on the space side, we look at it there. First, [S, S] is the poset of all order-preserving endofunctions on S. There are 22 = 4 functions from a two-element set to itself. In the case of a two-element chain, just one is not order preserving. That is, [S, S] = {0, idS, 1}, where 0 is the constantly 0 function, idS is the identity function, and 1 is the constantly 1 function. Since the projection map [S, S] → S must be surjective, 0 and 1 must be sent to 0 and 1, respectively. There

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

270

7 Domain Theory

are now two choices for where to send idS : for k ∈ {0, 1}, let p(k) denote the function [S, S]  S that sends 0 to 0, 1 to 1, and idS to k, see Figure 7.2. 1

1

idS

1

1

idS

0

0

0

p(0) [S, S]

0 p(1)

S

[S, S]

e (0)

S e (1)

Figure 7.2 The two embedding-projection pairs in Example 7.70. The solid lines represent the projections p(0) and p(1) and the dotted lines represent the embeddings e (0) and e (1) .

Both p(0) and p(1) have lower adjoints, e (0) and e (1) , respectively. Note that e (1) sends 0 to 0 and 1 to idS , while e (0) sends each element of S to the corresponding constant function. Note also that for each k ∈ {0, 1}, the pair (e (k), p(k) ) is an EPP, since all order-preserving maps between finite sets are spectral maps. We now choose to work with (e (0), p(0) ), because the function e (0) is the more natural embedding of S into [S, S]. Using Lemma 7.68, we obtain by induction a sequence of spaces Y0 := S

and

Yn+1 = [Yn, Yn ] for all n ≥ 0,

and a sequence of EPPs e0 := e (0), p0 := p(0)

and

en+1 := en ◦ ◦ pn, pn+1 := pn ◦ ◦ en

and thus we have a sequence of finite EPPs between domains S  [S, S]  [[S, S], [S, S]]  · · · and dually a sequence of finite EPPs between finite distributive lattices 3  FEnd (3)  FEnd (FEnd (3))  · · · . It remains to show that the inverse limit of the spaces, or equivalently, the colimit of the lattices is bifinite. This is easier to do on the lattice side, as we show now. n (3) for every n ≥ 0, denote by e : L  L Write L n := Fend n n n+1 : pn the EPP given above, and let L be the colimit of this sequence of EPPs in the category of distributive lattices, with en : L n → L, for each n ≥ 0, the colimit cone. As in the proof of Theorem 7.69, we use the fact that a directed colimit of inclusions is, up to isomorphism, just given by union, so that each en [L n ] is a sublattice of L. We

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

7.4 Domain Theory in Logical Form

271

claim that each en [L n ] is in fact an EPP-sublattice of L. The projection pn : L → L n is given for a ∈ L m by pn ◦ · · · ◦ pm−1 (a) if m > n, by the identity if m = n, and by en−1 ◦ · · · ◦ em if m < n. We leave it as an exercise to check that this is a well-defined embedding-projection pair (see Exercise 7.4.1). Thus, L is again bifinite. Now, FEnd (L) also has a colimit cone under the same diagram, which can be obtained by applying the functor FEnd to the colimit cone giving L. Therefore, Fend (L) is isomorphic to L. Example 7.71 We want to construct the least solution to the domain equation X  [X, X]⊥ , or dually, to the equation L  F→ (L, L)/ϑjpp ⊕ 1 within bifinite distributive lattices. The development is precisely the same as in the above example once we have an EPP to initialize the process. Again, it is simplest to argue on the space side. The terminal bifinite space is the one-element space, which we will denote by 1, and [1, 1]⊥ is the two-element ordered space ⊥ < 1. The difference here, which makes the process canonical, is that we have only one choice for p : [1, 1]⊥ → 1. Its lower adjoint is e : 1 → [1, 1]⊥, 1 → ⊥. Once we have this initial EPP, the arguments are identical to the ones in Example 7.70. We leave the details as an exercise. See Abramsky and Ong (1993, Section 4) for further details and the computational intuition behind the equation X  [X, X]⊥ . Exercises for Section 7.4 Exercise 7.4.1 Denote by Bif EPP the category of bifinite lattices with as morphisms the EPPs. Let U : Bif EPP → DL be the functor that is the identity on objects and sends a morphism (e, p) to e. Let D : I → Bif EPP be a directed diagram. Prove that the colimit of the diagram U ◦ D is again a bifinite lattice. Exercise 7.4.2 Fill in the details of the construction in Examples 7.70 and 7.71. Exercise 7.4.3 Let 2 ≤ω be the set of pairs (I, f ), where I is a down-set of the ordinal ω and f is a function I → 2. We put the strict prefix ordering on 2 ≤ω , that is, (I, f ) ≤ (J, g) if, and only if, I ⊆ J and g|I = f . Show that (2 ≤ω, ≤) is a bifinite domain. Notes for Chapter 7 The treatment of domain theory in this chapter focuses on this material as an application of Stone–Priestley duality and is based on this book’s authors’ interpretation of some of the material in Abramsky (1991). For more insight into denotational semantics and domain equations, we refer the interested reader to the specialized literature; we recommend in particular Abramsky and Jung (1994). Much more on λ-calculus and its models can be found, for example, in Barendregt (2014) or Barendregt and Manzonetto (2022).

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

272

7 Domain Theory

Abramsky (1987) was the first account, based on his PhD dissertation, of domain theory in logical form in the setting of Scott domains. Abramsky also already announced there the conceptual straightforwardness of extending to the bifinite setting modulo some technical intricacies to do with the identification of primes, and that this extension was being worked on by Glynn Winskel’s student Guo-Qiang Zhang at the time. Indeed, his PhD dissertation (Zhang, 1989), later published as Zhang (1991), takes, among others, Abramsky (1987) as starting point and treats the extension to the bifinite setting using the prevalent point of view in domain theory at the time. Concurrently, Abramsky (1991) also resolved the required intricacies based on an improvement of earlier work on bifinite domain theory in the dissertation of Gunter (1985). The Hoffmann–Lawson duality was established independently by Hoffmann (1981) and Lawson (1979). Note that this duality is not to be confused with the duality between distributive continuous lattices and locally quasi-compact sober spaces developed around the same time in Hofmann and Lawson (1978); the one for completely distributive complete lattices involved R. E. Hoffmann while the one for distributive continuous lattices involved K. H. Hofmann, and both dualities have the same J. D. Lawson as their second-named author. The results on completely distributive complete lattices presented in this chapter are due to Raney (1953), who introduced the relation ≪ many years before the introduction of the relation ) in Scott (1972). Thus, it is somewhat backward to introduce ≪ as a more specialized version of ) as, given the chronology, ) was in fact a generalization of the already existing ≪.

https://doi.org/10.1017/9781009349680.008 Published online by Cambridge University Press

8 Automata Theory

In this chapter, we discuss applications of duality theory to automata theory. The theory of regular languages and automata is a computer science topic with a rich theory and a wide and still growing range of practical and theoretical applications (Almeida et al., 2020; Pin, 2021a, 2021b). The availability of sophisticated mathematical tools from algebra and topology is one of the main strengths of the classical framework. Finite algebras were introduced into the theory early on by Myhill, and Rabin and Scott, and their power was established by the effective characterization of star-free languages by means of syntactic monoids (Schützenberger, 1955, 1965). Syntactic monoids provide an abstract and canonical notion of so-called recognition for regular languages. Eilenberg’s theorem (Eilenberg, 1976) supplies a general framework in which to apply the strategy of Schützenberger’s result. The success of this so-called algebraic method was greatly augmented by the introduction in the 1980s of profinite topology and in particular profinite monoids, see Almeida (1995, 2005), Weil (2002), and Pin (2009). A most powerful combination in this setting is that of Eilenberg’s and Reiterman’s theorems. Reiterman’s theorem is a generalization of Birkhoff’s variety theorem from universal algebra. It states that pseudovarieties of finite algebras are precisely the ones given by profinite identities (Reiterman, 1982). Thus, Eilenberg–Reiterman theory allows for the equational description of certain classes of regular languages. This can lead to decidable criteria for membership in the corresponding classes, in case a finite equational basis can be found, or, at least, an infinite basis that can be effectively tested on a finite algebra (Almeida and Weil, 1998). In this chapter, we will see the rudiments of this theory from a duality point of view. Thus, recognition by monoids, and in particular syntactic monoids, will be defined duality-theoretically. We will see that the dual space of the Boolean algebra of regular languages over a given alphabet with certain additional operations is the so-called free profinite monoid. We will then show that the subalgebra-quotient

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

274

8 Automata Theory

duality of Section 4.2, when applied to this pair, leads to the profinite equational method of Eilenberg–Reiterman theory. We will not cover the Eilenberg–Reiterman theory in full generality, but will see a particular example of it and we provide references to the recent research literature throughout the chapter for readers who want to get the full general picture. More bibliographic references can also be found in the notes at the end of this chapter. In Section 8.1, after introducing basic notions from automata theory, we show how the classical construction of a syntactic monoid for a regular language can be understood via duality. In particular, we will see how this is a special case of the general duality for implication operators developed earlier in Section 4.5. The results in the first section use discrete duality in the case of free monoids and finite monoids, from which we then also derive that the collection of all regular languages is a Boolean algebra. Section 8.2 shows how to extend these ideas to profinite monoids, and establishes as its main result that the free profinite monoid is the dual of the Boolean algebra of regular languages, enriched with some additional operations. We show also how this result fits in a more general duality theorem for Boolean spaces equipped with a binary operation. Sections 8.3 and 8.4 show how this allows one to apply the methods of duality theory in the setting of regular languages, focusing in particular on characterizations of classes of languages via profinite equations. A more detailed outline of the rest of the chapter is given at the beginning of Section 8.2, after we introduce some of the basic notions that we will need. 8.1 The Syntactic Monoid as a Dual Space We started our introduction to duality theory back in Chapter 1 with the discrete duality for finite distributive lattices, see Section 1.3. The finite case came first chronologically and, while it is simpler than the full topological theory, it contains many of the central ideas of the subject in embryonic form. A foundational result of algebraic automata theory is the fact that one may associate, with each finite state automaton, a finite monoid with a universal property. Our first goal in this chapter is to show that this foundational result is a case of discrete duality for complete and atomic Boolean algebras with additional operations. This will then lead to the applications of duality theory to the theory of regular languages that we look at later in the chapter. Automata, Quotienting Operations, and Residuation We start with automata, and show how they naturally lead us to consider dual Boolean algebras equipped with quotienting operations. An automaton is a very

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

275

simple and weak model of computation; it is a machine that reads once over a finite sequence of input values and only in the order in which it is entered, and only produces one output bit at the end of its computation. Readers who have previously encountered notions of computability will notice that automata are very far from being Turing complete (see, for example, Exercise 8.1.3(e)). We will typically denote the input alphabet by A and will assume throughout that it is finite. In fact, profinite input alphabets are also natural in this setting, but we shall not need them in what we discuss in this chapter. Recall that a monoid is a triple (M, ·, 1), where · is an associative binary operation on M and 1 is a neutral or identity element for ·, that is, 1 · m = m = m · 1 for all m ∈ M. We will often omit · in the notation, writing simply mn instead of m · n, as is common practice in algebra. When multiple monoids M and N are in play, we sometimes use the notations 1 M and 1 N to distinguish their neutral elements. Note that, for an input alphabet A, the set of all finite sequences over A with binary concatenation is the free monoid generated by A, see Example 5.17(e) and Exercise 5.2.10. Recall that we denote this algebra by A∗ . We suppress the commas in sequences, so that we think of an element w ∈ A∗ as a finite A-word, often simply called a word. That is, a word is a possibly empty string w = a1 · · · an , where n ≥ 0 is a number called the length of w and ai ∈ A for 1 ≤ i ≤ n. Notation If w ∈ A∗ is a finite word, as a notational convention, whenever we write w = a1 · · · am , we mean that the ai are elements of the alphabet A, and thus that m is equal to the length of w. We now define the notion of automaton, also known as non-deterministic finite state automaton in the literature. Definition 8.1 A (finite state) automaton is a structure A = (Q, A, δ, I, F), where (a) (b) (c) (d) (e)

Q is a finite set, whose elements we call states; A is a finite set, called the (input) alphabet; δ is a subset of Q × A × Q, called the transition relation; I ⊆ Q is a set, whose elements we call initial states; and F ⊆ Q is a set, whose elements we call final or accepting states.

We will call a tuple of the form M = (Q, A, δ) satisfying (a)–(c), without the initial and finite states specified, a (finite state) machine. Given an automaton A = (Q, A, δ, I, F), we call the associated machine M = (Q, A, δ) the underlying machine of A. Let w = a1 · · · an ∈ A∗ with each ai ∈ A. An automaton A accepts w if there exists a sequence of states q0, . . . , qn ∈ Q such that q0 ∈ I, qn ∈ F and (qi−1, ai, qi ) ∈ δ for every 1 ≤ i ≤ n; if no such sequence exists, A rejects w.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

276

8 Automata Theory

In particular, note that A accepts the empty word  if, and only if, I ∩ F  ∅. We denote by L(A) the set of words accepted by A, which is also known as the language recognized by A. A subset L ⊆ A∗ is called a regular language if there exists an automaton A such that L = L(A). 1 The above definition allows for non-determinism: When in state q, upon reading a letter a ∈ A, the automaton has the possibility to move to any of the states in the set δ[q, a, _]. When this set is always non-empty, the automaton is called complete, and when this set has cardinality at most 1, and the set of initial states has cardinality 1, the automaton is called deterministic. 2 Any regular language of finite words can be recognized by a complete deterministic automaton, as will also follow from the characterizations we give in Theorem 8.22. We note, however, that determinism does not come for free: Exercise 8.1.2 gives an example of a non-deterministic automaton with n states, but for which any deterministic automaton recognizing the same language requires at least 2n states. The connection with duality that we expose here is natural for the general notion of finite state machine that we defined above, and does not require determinism a priori. When M = (Q, A, δ) is a machine, for each a ∈ A, we will write δ a := {(q, q  ) ∈ Q × Q | (q, a, q  ) ∈ δ}, which is a binary relation on Q. We extend this definition to words as follows. For any non-empty word w = a1 · · · am ∈ A∗ , let δ x be the relational composition δ a1 · · · · · δ am and also define δ to be the identity relation on Q. A different way to say this is that w → δw is the unique monoid homomorphism from the free monoid A∗ to the monoid of binary relations on Q extending the assignment a → δ a , see Exercise 8.1.9. The relation δw is the reachability relation of the machine M when reading a word w ∈ A∗ . A word w is then accepted by an automaton A = (Q, A, δ, I, F) based on M if, and only if, there exist q0 ∈ I and q f ∈ F such that (q0, q f ) ∈ δw . Example 8.2 Consider the deterministic finite-state automaton A = (Q, A, δ, {1}, {2}), where Q := {1, 2, 3}, A := {a, b}, and δ := {(1, a, 2), (2, b, 3), (3, a, 2)}. We will often depict such an automata by a graph, as in the following diagram. 1

2

The name regular language originates from a different approach, via regular expressions, that is, expressions, built up from single letters by finite union, finite concatenation product, and the Kleene star, which yields the submonoid of A∗ generated by a subset of A∗ . The concept we have defined is more aptly called recognizable language. However, it is a non-trivial theorem, proved by Kleene at the very beginning of automata theory, that a language is recognizable if, and only if, it is regular (Kleene, 1956), and the adjective regular is more commonly used, even if one focuses on recognition as we do here. In some references, the term “deterministic” includes “complete” by definition. Also, note that we can safely ignore the case where the set of initial states I is empty.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

277

b 1

a

3

2 a

Here, the nodes of the graph are the states and the directed edges are labeled by letters and correspond to the elements of the transition relation. The initial states are indicated by an entering arrow while the final states are indicated by an arrow out. From the definition of acceptance, note that a word w is accepted by this automaton A if, and only if, w begins with an a, followed by an arbitrary number of repetitions of the string ba. Thus, we have our first example of a regular language, L(A) = {a(ba) n | n ≥ 0} = {a, aba, ababa, abababa, . . . }. The same transition relation δ on this set of states Q may be used to recognize other regular languages; for example, the reader may verify that the automaton A  := (Q, A, δ, {2}, {2, 3}) recognizes the language L(A  ) = {(ba) n, (ba) n b | n ≥ 0}. Other regular languages can be obtained by choosing other sets of initial and final states. Crucially, however, if we fix any finite state machine M = (Q, A, δ), then there are only finitely many choices for a pair of subsets (I, F) ∈ P (Q) 2 . Thus, only finitely many regular languages can be recognized by the automata based on a fixed machine M. This essentially trivial observation will be important in what follows. We will build on the idea of the above example to give an algebraic characterization of regular languages as certain special elements of the Boolean algebra P ( A∗ ), enriched with additional operations. Definition 8.3 Let A = (Q, A, δ, I0, F0 ) be an automaton and let L = L(A) and M = (Q, A, δ) the underlying machine of A. For any pair of subsets (I, F) ∈ P (Q) 2 , denote by L(I, F) the language recognized by the automaton (Q, A, δ, I, F); in particular, L(A) := L(I0, F0 ). We write R (M) := {L(I, F) | (I, F) ∈ P (Q) 2 } ⊆ P ( A∗ ) for the set of languages which can be recognized by automata obtained from A by changing its initial and final states. We denote by L(M) the sublattice generated by R (M) and by B(M) the Boolean subalgebra generated by R (M). Remark 8.4 Let M be a finite state machine. Note that the set R (M) has at most 22 |Q | elements and is thus in particular finite. Therefore, L(M) and B(M) are both finite, since finitely generated distributive lattices and Boolean algebras are finite

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

278

8 Automata Theory

(see Exercise 4.1.3). It follows that both L(M) and B(M) are, trivially, complete sublattices of P ( A∗ ). The idea that we will develop now is that the finite lattice L(M) and the finite Boolean algebra B(M) can be viewed as dual algebraic structures associated to the machine M. The dynamic behavior of a machine M is captured by the action of the alphabet on the set of states. Taking into account the direction-reversing nature of duality, it stands to reason that the appropriate structure to consider on the dual collection of languages is given by quotienting operations, defined, for any language L ⊆ A∗ and letter a ∈ A, by a−1 L := {w ∈ A∗ | aw ∈ L}

and

La−1 := {w ∈ A∗ | wa ∈ L},

or more generally, for any word x ∈ A∗ , x −1 L := {w ∈ A∗ | xw ∈ L},

Lx −1 := {w ∈ A∗ | wx ∈ L}.

We call the operations L → x −1 L and L → Lx −1 the left and right quotienting operations on the Boolean algebra of languages P ( A∗ ). Note that these operations are Boolean algebra endomorphisms on P ( A∗ ); indeed, each of the operations computes the inverse image of a function from A∗ to A∗ . These quotienting operations may thus be seen as a biaction of the free monoid A∗ on the Boolean algebra P ( A∗ ); we will further explore this in Definition 8.13. For now, we say that a collection of languages L ⊆ P ( A∗ ) is closed under quotienting if, for every x ∈ A∗ and L ∈ L, we have that x −1 L ∈ L and Lx −1 ∈ L. A key insight is that the collection R (M) of languages recognized by a machine M is closed under quotienting, as we will prove now. Recall that, when R is a relation from a set X to Y , we write R[U] for the relational direct image of U under R, that is, R[U] = {y ∈ Y | uRy for some u ∈ U}, and similarly R−1 [V ] for the relational inverse image. We apply this notation here to the reachability relations δ x for x ∈ A∗ induced by a finite state machine. Lemma 8.5 Let M = (Q, A, δ) be a machine. For any x, y ∈ A∗ and (I, F) ∈ P (Q) 2 , we have x −1 L(I, F) = L(δ x [I], F)

and

L(I, F)y −1 = L(I, δ−1 y [F]).

In particular, R (M) is closed under quotienting. Proof Note that (q, q  ) ∈ δ x if, and only if, there exists a path from q to q  labeled by the letters of the word x. Using this characterization, for any x, w ∈ A∗ and (I, F) ∈ P (Q) 2 , we now have that w ∈ x −1 L(I, F) if, and only if, xw ∈ L(I, F),

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

279

if, and only if, there exist states q0, q1, q2 ∈ Q such that q0 ∈ I, (q0, q1 ) ∈ δ x, (q1, q2 ) ∈ δw and q2 ∈ F. This last condition is clearly equivalent to the condition that the word w is in the language L(δ x [I], F). This proves the first equality. The proof of the second equality is similar.  We have thus associated, with any finite state machine M, the set R (M) which, as Lemma 8.5 shows, is closed under the quotienting operations. We now show that it follows from this that the lattice L(M) and the Boolean algebra B(M) are also closed under quotienting, which means that they can be considered as algebraic structures in and of themselves, independently from the machine M. Proposition 8.6 Let M = (Q, A, δ) be a machine. Then, L(M) and B(M) are closed under quotienting. Proof Let x ∈ A∗ . We will prove that x −1 L ∈ B(M) for every L ∈ B(M); the proofs of the other properties are essentially the same. Let Bx denote the inverse image of B(M) under the quotienting operation by x, that is, Bx := {L ∈ P ( A∗ ) | x −1 L ∈ B(M)}. Then, since the quotienting operation is an endomorphism, Bx is a Boolean subalgebra of P ( A∗ ) and it contains R (M) by Lemma 8.5. Thus, Bx contains B(M), which is what we needed to prove.  Combining Remark 8.4 and Proposition 8.6, we now aim to deduce stronger closure properties for L(M) and B(M), namely, they are not just closed under quotienting by single words, but by any set of words. In order to express this closure property, which also plays a central role in the more general theory of varieties of regular languages alluded to in the introduction of this chapter, we need to introduce residuation operations on P (M) for an arbitrary monoid M. For any monoid M, the Boolean algebra P (M) carries a complex multiplication · defined by u · t := { mn | m ∈ u, n ∈ t }. Since this operation · is a complete binary operator on a complete Boolean algebra, it has a left and a right residual, which we denote by \ and /, respectively, see Example 4.62 and Exercise 4.5.6. These are the two unique operations, \, / : P (M) 2 → P (M), such that, for any t, u, v ∈ P (M), u · t ⊆ v ⇐⇒ t ⊆ u\v ⇐⇒ u ⊆ v/t. More explicitly, in terms of our duality for implication-type operators of Section 4.5, let us write R ⊆ M 3 for the graph of the multiplication on M, that is,

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

280

8 Automata Theory

R = {(m1, m2, m) ∈ M 3 | m1 · m2 = m}. Then, \ and / are the implication-type operators obtained from the relation R given by the monoid multiplication on M by using the first and the third, or, respectively, the second and the third coordinates as inputs: u\v = M − (R[u, _, M − v]) = {m ∈ M | km ∈ v for all k ∈ u}, v/t = M − (R[_, t, M − v]) = {m ∈ M | m ∈ v for all  ∈ t}. In fact, as we will see below, if we want to witness a monoid operation · on a set M on its discrete dual Boolean algebra P (M), then the residual operations \ and / are more important than the complex multiplication. This is not so surprising since duality should reverse the direction of a map. Note that, starting with the following definition, we distinguish between residuated algebras, where the multiplication · is present as part of the structure, and residuation algebras, where the multiplication · is absent. While the difference is small in this section, we will later consider residuation algebras in more generality, see Definition 8.27. Definition 8.7 Let M be a monoid. We call the Boolean algebra with three additional operations (P (M), ·, \, /) the (discrete) dual residuated Boolean algebra of M. The reduct (P (M), \, /), a Boolean algebra with two additional operations, will be called the dual residuation algebra of M. A residuation ideal is a sublattice L of P (M) such that, for any r ∈ P (M) and s ∈ L, both r\s ∈ L and s/r ∈ L. A residuation ideal is said to be Boolean provided the underlying sublattice is a Boolean subalgebra and it is said to be complete provided the underlying sublattice is a complete sublattice of P (M). Residuation ideals are closed under intersections, so that for any collection S ⊆ P (M), there is a smallest residuation ideal containing S, which we call the residuation ideal generated by S. Note also that the collections of Boolean, complete, and complete Boolean residuation ideals are also all closed under intersections and thus there are also corresponding notions of generated complete and/or Boolean residuation ideals (see Exercise 8.1.6 and Definition 8.11). Quotienting operations may now be defined for a general monoid M by restricting the residual operations to the atoms of P (M) as follows. Definition 8.8 Let M be a monoid, m ∈ M and s ∈ P (M). We write m−1 s := {m}\s

and

sm−1 := s/{m}.

We call m−1 s and sm−1 the left and right quotients of s by m, respectively. We will call the operations s → m−1 s and s → sm−1 quotienting operations on the Boolean algebra P (M). A subset S ⊆ P (M) is closed under quotienting if, for every s ∈ S and m ∈ M, m−1 s ∈ S and sm−1 ∈ S. The closure under quotienting of a subset S ⊆ P (M) is

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

281

defined as the smallest subset Q(S) that contains S and is closed under quotienting; concretely, Q(S) = {m−1 sn−1 | m, n ∈ M, s ∈ S}. Note in particular that the operations L → x −1 L and L → Ly −1 , for x, y ∈ A∗ and L ⊆ A∗ , which we introduced in Lemma 8.5, are the quotienting operations on P ( A∗ ) given by the monoid A∗ . Being closed under quotienting is clearly a consequence of being a residuation ideal; we now prove that the converse holds in the complete setting. Proposition 8.9 Let M be a monoid and let L be a complete sublattice of P (M). Then, L is a residuation ideal if, and only if, L is closed under quotienting. Proof Suppose L is a complete sublattice of P (M) that is closed under quotient ing. Let s ∈ L and r ∈ P (M). Since r = m∈r {m}, we have  r\s =  {m} \s = m−1 s, m∈r m∈r  and the latter is in L because L is closed under quotienting and complete. In the same way, s/r ∈ L.  Summarizing what we know about the lattice L(M) and the Boolean algebra B(M) associated to a finite state machine M, we have the following. Proposition 8.10 Let M = (Q, A, δ) be any finite state machine. Then, L(M) and B(M) are finite residuation ideals of P ( A∗ ). Proof By Proposition 8.6, L(M) and B(M) are closed under quotienting. By Remark 8.4, L(M) and B(M) are finite, hence complete, sublattices of P ( A∗ ). Proposition 8.9 now gives the statement.  Proposition 8.10 will allow us to express regularity of a language L in a way that is independent from any particular machine recognizing L, only using the residuation structure on P ( A∗ ). Definition 8.11 For any language L ∈ P ( A∗ ), we write L(L) for the residuation ideal of P ( A∗ ) generated by {L} and B(L) for the Boolean residuation ideal of P ( A∗ ) generated by {L}. Corollary 8.12 Let L ∈ P ( A∗ ) be regular. Then, L(L) and B(L) are finite and thus complete residuation ideals of P ( A∗ ). Moreover, both are effectively computable from any automaton that recognizes L.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

282

8 Automata Theory

Proof Let A be any automaton that recognizes L and write M for its underlying machine. Then, L ∈ L(M), so by Proposition 8.10, we must have L(L) ⊆ L(M), and the latter is finite. In particular, L(L) can be computed effectively by closing {L} under quotienting and lattice operations of L(M), which are given by Lemma 8.5. The argument for B(L) is the same.  We will see in Theorem 8.22 that the conditions in Corollary 8.12 are in fact enough to conclude that L is regular. For now, we see that L(L) and B(L) are finite algebraic structures canonically associated with a regular language L. We will show in the next subsection that B(L) is in fact the Boolean residuation algebra dual to the syntactic monoid of the regular language L.

Discrete Duality for Monoid Quotients and Residuation Ideals We will now need to use some universal algebraic notions for monoids. The general definitions of universal algebra apply to monoids just as they apply to lattices or Boolean algebras; we briefly recall the basics of these definitions in the case of monoids and refer to, for example, Burris and Sankappanavar (2000) or Wechler (1992) for more information. A function f : M → N between monoids is called a homomorphism if f (xy) = f (x) f (y) for all x, y ∈ M, and f (1 M ) = 1 N . A congruence on a monoid M is an equivalence relation ϑ such that, for any m1, m1 , m2, m2 ∈ M, if mi ϑ mi for i = 1, 2, then m1 m2 ϑ m1 m2 . Note that it is equivalent to say that ϑ is invariant under multiplication on both sides, that is, if m, m  ∈ M with m ϑ m  and x ∈ M, then xm ϑ xm  and mx ϑ m  x. The quotient of M by the congruence ϑ is the monoid based on the set of ϑ-equivalence classes, M/ϑ, uniquely defined by the requirement that [m1 ]ϑ [m2 ]ϑ = [m1 m2 ]ϑ for every m1, m2 ∈ M; its neutral element is [1 M ]ϑ . For any homomorphism f : M → N between monoids, the kernel of f , ker( f ) := {(m, m  ) | f (m) = f (m  )}, is a congruence on M and f can be factored as the surjective homomorphism M  M/ker( f ) that sends m to [m]ker( f ) followed by an injective homomorphism M/ker f → N well defined by requiring that [m]ker( f ) is sent to f (m), for every m ∈ M. This construction is known as the first isomorphism theorem for monoids. In particular, for any surjective monoid homomorphism M  N, N is isomorphic to M/ker( f ), so that surjective homomorphisms with domain M may be identified, up to isomorphism, with congruences on M. As in the case of distributive lattices (see Exercise 5.3.1), surjective homomorphisms in the category of monoids are epimorphisms, but the converse is not true (see Exercise 8.1.8).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

283

A semigroup is a pair (S, ·) where · is an associative binary operation on M. Thus, “semigroup = monoid minus neutral element.” 3 A subsemigroup of a monoid M is a subset of M that is closed under multiplication. A subsemigroup is a submonoid if it moreover contains 1. Notice that a subsemigroup of a monoid M may be a monoid with a different identity element than M (see Exercise 8.1.5). If M is a monoid and S is a subset of M, then there exist a smallest subsemigroup and a smallest submonoid of M that contain S. These are called the subsemigroup generated by S and the submonoid generated by S. A cornerstone result in the algebraic theory of regular languages establishes a close connection between languages recognized by finite automata and finite quotients of free monoids. Our aim in the rest of this section is to show that this connection is an instance of discrete duality. We saw in Section 4.2, Theorem 4.36, that quotients of a set correspond to complete Boolean subalgebras of its power set. We are now in a situation where the set M additionally carries the structure of a monoid, and we will show in Theorem 8.18 that the monoid quotients of M correspond dually to complete Boolean residuation ideals of P (M). The first step toward our proof of Theorem 8.18 will be to view monoid quotients as biactions; see Proposition 8.17. We will then use this theorem to characterize regular languages. We first give the necessary algebraic definitions. Definition 8.13 Let M be a monoid and S a set. A left action of M on S is a function λ : M × S → S such that, for every m, m  ∈ M and x ∈ S, we have λ(m  m, x) = λ(m , λ(m, x)) and λ(1 M , x) = x. A left action may also be viewed as a homomorphism from M to the monoid of functions S → S under composition, by sending m ∈ M to the function λ m : S → S defined by λ m (x) := λ(m, x); the left action is called faithful if m → λ m is injective. If the left action is clear from the context, we write m · x for λ(m, x), noting that the action law then gives that (m  m) · x = m  · (m · x). If M has left actions on two sets S and S , then a function h : S → S  is a morphism if, for any m ∈ M, x ∈ S, we have m ·S h(x) = h(m ·S x). The notion of a right action of M on S is defined symmetrically. A biaction of M on S is a pair (λ, ρ), where λ is a left action, ρ is a right action, and, for any m, m  ∈ M and x ∈ S, we have ρ(λ(m, x), m  ) = λ(m, ρ(x, m  )). This axiom implies that, when the pair (λ, ρ) is clear from the context, we can unambiguously write m · x · m  for the result of acting on x by m on the left and 3

We will not be concerned much with semigroups in this chapter, although most of the results that we state and prove for monoids also hold for semigroups.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

284

8 Automata Theory

by m  on the right, with the actions being performed in whichever order. A biaction can thus be described alternatively as a function α : M × S × M → S for which α(−, −, 1 M ) is a left action, α(1 M , −, −) is a right action, and α(m, x, m  ) = α(1 M , α(m, x, 1 M ), m  ) = α(m, α(1 M , x, m  ), 1 M ) for every m, m  ∈ M and x ∈ S. A biaction morphism is a function that is an action morphism for both the induced left and right actions. Example 8.14 The Cayley representation of a monoid M is given by letting it act on its own underlying set: for any m, x ∈ M, define λ(m, x) := mx. This left action is faithful, since λ m (1 M ) = λ m (1 M ) implies m = m . Similarly, the monoid M acts on itself on the right, by setting ρ(x, m) := xm for any x, m ∈ M. The associativity axiom of the monoid says that (λ, ρ) is a biaction, and we call this the Cayley biaction of M on itself. Applying discrete duality to the Cayley functions λ m and ρm , for each m ∈ M, −1 we obtain two complete homomorphisms λ −1 m , ρ m : P (M) → P (M). These two functions then give a biaction of M on the set P (M), concretely given by: M × P (M) × M → P (M), (m, s, n) → {x ∈ M | mxn ∈ s}. This biaction of M on P (M) corresponds to looking at residuation with respect to singleton sets. As long as we stick to the discrete duality, it is equivalent whether one considers residuation by general sets or the biaction given by residuation with respect to singleton sets. This fact is exploited in the development below, in particular in the proof of Theorem 8.18. We now show that biactions, taken through discrete duality, give biactions. Proposition 8.15 Let M be a monoid and suppose M has a biaction on S. Then, M has a biaction on P (S) given, for m1, m2 ∈ M and u ∈ P (S), by m1−1 u m2−1 := {x ∈ S | m1 · x · m2 ∈ u}, and the map u → m1−1 um2−1 is a complete Boolean algebra homomorphism. The induced left action is u → um2−1 and the induced right action is u → m1−1 u. If M also has a biaction on T and h : S → T is a biaction morphism, then h−1 : P (T ) → P (S) is a biaction morphism for the dual biactions. In particular, for any m1, m2 ∈ M, the quotienting operation P (M) → P (M) given by u → m1−1 u m2−1 is a complete Boolean algebra homomorphism. Proof The fact that (m, x) → m· x is a left action of M on S corresponds via duality to the fact that (u, m) → m−1 u is a right action of M on P (S), and symmetrically on the other side. Note that for this reason we write the dual of the right action on the left and the dual of the left action on the right. The compatibility of the two actions on S

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

285

is dual to the compatibility of the two actions on P (S), and likewise the dual h−1 of h is a biaction morphism (see Exercise 8.1.7 for details). The “in particular” statement is the application of the proposition to the Cayley biaction of M on itself.  We note a consequence of Proposition 8.15, which generalizes the argument given in Proposition 8.6. Recall from Definition 8.8 that, when M is a monoid and S is a subset of P (M), Q(S) denotes the closure under quotienting of S. Proposition 8.16 Let M be a monoid and S ⊆ P (M). The (complete) sublattice generated by Q(S) and the (complete) Boolean subalgebra generated by Q(S) are all closed under quotienting. In particular, for any regular language L, L(L) is the sublattice of P ( A∗ ) generated by Q({L}) and B(L) is the Boolean subalgebra of P ( A∗ ) generated by Q({L}). Proof Denote by L the sublattice generated by Q(S). Let m1, m2 ∈ M be arbitrary. Since the function h : u → m1−1 um2−1 is a lattice homomorphism, the inverse image h−1 (L) is a sublattice of P (M). Also, note that Q(S) ⊆ h−1 (Q(S)) ⊆ h−1 (L), and thus L ⊆ h−1 (L). We conclude that, for any u ∈ L, m1−1 um2−1 ∈ L. Thus, L is closed under quotienting. The proofs for the generated complete sublattice, the generated Boolean subalgebra, and the generated complete Boolean subalgebra are analogous. The “in particular” statement now follows by the proof of Corollary 8.12.  In the following proposition, we will consider a monoid M, a set S, and a surjective function h : M  S from M onto a set S. Note that, since h is surjective, there is at most one binary operation ·S on S with respect to which h becomes a homomorphism, namely, whenever s1, s2 ∈ S, we can pick mi ∈ M such that si = h(mi ), and then s1 ·S s2 must be h(m1 m2 ). This operation ·S is well defined if, and only if, the kernel of h is a monoid congruence. We will now show that this is also equivalent to having a biaction of M on S that is respected by the function h. Proposition 8.17 Let M be a monoid, S a set, and let h : M  S be a surjective function. The following are equivalent: (i) (ii)

there exists a well-defined monoid structure on S such that h is a homomorphism; and there exists a biaction of M on S such that h is a morphism from the Cayley biaction on M to the biaction on S.

Proof Assume (i) holds. Define the biaction of M on S by αh : M × S × M → S

and

α(m1, s, m2 ) := h(m1 ) ·S s ·S h(m2 ).

(8.1)

We first show that (m, s) → αh (m, s, 1 M ) is a left action. Indeed, for any m, m  ∈ M and s ∈ S we have

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

286

8 Automata Theory

αh (m  m, s, 1 M ) = h(m  )h(m)s = αh (m , αh (m, s, 1 M ), 1 M ) and also, αh (1 M , s, 1 M ) = h(1 M )sh(1 M ) = s. Symmetrically, (s, m) → αh (1 M , s, m) is a right action. By a similar calculation, αh is a biaction. Finally, h is a morphism of biactions, since, for the left actions, for any m, m  ∈ M, we have h(m · M m  ) = h(mm  ) = h(m)h(m  ) = αh (m, h(m  ), 1 M ), and similarly for the right action. Now assume (ii) holds and let (λ, ρ) be such a biaction. Then, for any m1, m2 ∈ M, we have h(m1 m2 ) = h(λ m1 (m2 )) = α(m1, h(m2 ), 1 M ), and similarly h(m1 m2 ) = α(1 M , h(m1 ), m2 ). From this, it follows that if m1 ∈ M and (m2, m2 ) ∈ ker(h), then we must have h(m1 m2 ) = α(m1, h(m2 ), 1 M ) = h(m1 m2 ) and, symmetrically, we also have h(m2 m1 ) = h(m2 m1 ). Thus, ker(h) is a monoid congruence, from which (i) follows by the remarks preceding the proposition.  Propositions 8.17 and 8.15 together imply that surjective monoid homomorphisms M  N are dual to maps P (N ) → P (M) that preserve the quotienting operations on the power set algebras of M and N. We now characterize the images of such maps as the complete Boolean residuation ideals of P (M). Theorem 8.18 Let M be a monoid and (P (M), ·, \, /) its discrete dual residuated Boolean algebra. Let h : M  N be a surjective function from M to a monoid N and denote by B the image of h−1 : P (N ) → P (M). Then, h is a monoid homomorphism if, and only if, B is a residuation ideal of P (M). In particular, under discrete duality, monoid quotients of M correspond to complete Boolean residuation ideals of P (M). Proof First suppose h is a monoid homomorphism. By Proposition 8.17, h induces a biaction of M on N and by Proposition 8.15, h−1 is a morphism for the dual biaction of M on P (N ), and thus the image of h−1 is closed under quotienting. By Proposition 8.9, it is therefore a residuation ideal. For the converse, suppose that the image B of h−1 is a residuation ideal. Then, in particular we get a biaction of M on P (N ), as follows. For any m, m  ∈ M and u ∈ P (N ), let m · u · m  be the unique element of P (N ) such that h−1 (m · u · m  ) = m−1 h−1 (u)(m  ) −1 .

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

287

Note that the latter is indeed in B because it is a residuation ideal. The dual of this biaction of M on P (N ) now gives a biaction of M on N for which h is a biaction morphism (see Exercise 8.1.7), so Proposition 8.17 implies that h is a monoid homomorphism.  In summary, Theorem 8.18 shows that the anti-isomorphism of Theorem 4.36, between equivalence relations on a set M and complete Boolean subalgebras of P (M), restricts to an anti-isomorphism between monoid congruences on M and complete Boolean residuation ideals of P (M). For any surjective monoid homomorphism h : M  N, we define the corresponding residuation ideal of P (M) to be B := {h−1 (u) | u ⊆ N }. Conversely, when B is a complete Boolean residuation ideal of P (M), with M a monoid, we have a unique congruence ϑ on M for which B is the corresponding residuation ideal; we call ϑ the congruence dual to B. Recall from the proof of Theorem 4.36 that this congruence ϑ may be defined explicitly, for x, y ∈ M, by xϑy ⇐⇒ for all a ∈ B, x ∈ a if, and only if, y ∈ a. The following definition, namely that of recognition from automata theory, codifies a slight generalization of the duality relationship of Theorem 8.18, in that h need not be surjective. Definition 8.19 Let M be a monoid, s ∈ P (M) and h : M → N a monoid homomorphism. We say that s is recognized by h provided there is u ⊆ N such that s = h−1 (u); in this case, we call h a recognizer for s. That is, s is recognized by h if, and only if, it belongs to the residuation ideal corresponding to the surjective homomorphism M  im(h). Further, we say that a monoid N recognizes s provided there is a monoid homomorphism h : M → N which recognizes s. When h : M → N is a monoid homomorphism, we call the surjective function h : M  im(h), which is still a homomorphism, now onto the submonoid im(h) of N, the surjective co-restriction of the homomorphism h. It follows from Theorem 8.18, combined with Theorem 4.36, that we always have a least recognizer for any s ∈ P (M), namely the dual of the complete Boolean residuation ideal generated by s. Corollary 8.20 Let M be a monoid and let s ∈ P (M). Then, the complete Boolean residuation ideal generated by s corresponds to the monoid quotient hs : M  M/≡s,

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

288

8 Automata Theory

where m ≡s m  if, and only if, (xmy ∈ s ⇐⇒ xm  y ∈ s).

for all x, y ∈ M

This is the least recognizer of s, in the sense that it factors through the surjective co-restriction of any other recognizer. The first statement of Corollary 8.20 follows from the fact that the complete Boolean residuation ideal generated by s is equal to the smallest complete Boolean subalgebra containing Q({s}), by Propositions 8.9 and 8.16. We leave the details of the rest of the Corollary as Exercise 8.1.10. Definition 8.21 Let M be a monoid and s ∈ P (M). Then, the congruence ≡s , defined by m ≡s m  if, and only if, for all x, y ∈ M,

xmy ∈ s ⇐⇒ xm  y ∈ s

is called the syntactic congruence of s. The quotient map hs : M  M/≡s is called the syntactic homomorphism of s and Ms = M/ ≡s is called the syntactic monoid of s. We are now ready to prove the main theorem of the section. The equivalence of (i), (iv), and (v) in Theorem 8.22 is a classical result in the automata theory literature, but our point of view here, via duality for residuation ideals, is more recent (Gehrke et al., 2008). Theorem 8.22 Let L ∈ P ( A∗ ). The following conditions are equivalent: (i) (ii) (iii) (iv) (v)

the language L is regular; the residuation ideal L(L) is finite; the Boolean residuation ideal B(L) is finite; the syntactic monoid ML of L is finite; and the language L is recognized by a finite monoid.

Moreover, if L is regular then the syntactic monoid is effectively computable from any automaton recognizing L. Proof If L is regular, then, by Corollary 8.12, L(L) is finite. If L(L) is finite, then B(L) is finite as it is the Boolean algebra generated by L(L). If B(L) is finite, then its dual set is finite, and this is the syntactic monoid ML , by Corollary 8.20. Note that this dual of B(L) is indeed effectively computable, since B(L) can be computed from any automaton recognizing L by Corollary 8.12, and its dual is the set of atoms of B(L) equipped with the multiplication dual to the quotienting

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

289

operations. If ML is finite, then as L is recognized by ML , it is recognized by a finite monoid. Finally, if L is recognized by a finite monoid N, then there is a monoid homomorphism h : A∗ → N and u ⊆ N so that L = h−1 (u). We define an automaton A = (N, A, δ, {1 N }, u) by setting (n, a, n  ) ∈ δ

nh(a) = n  .

⇐⇒

Then, for x = a1 · · · an ∈ A∗ , we have that A accepts x if, and only if, h(x) =  h(a1 ) · · · h(an ) = 1 N h(a1 ) · · · h(an ) ∈ u if, and only if, x ∈ L. Definition 8.23 Let A be a finite set. We denote by Reg A the subset of P ( A∗ ) consisting of the regular languages over A. Using the characterization of regular languages in Theorem 8.22, we can now show that the set Reg A of all regular languages in fact carries a lot of algebraic structure. Apart from its interest in computer science, the infinite Boolean algebra Reg A is of interest in constructive mathematics as it is a non-trivial countably infinite subalgebra of a power set algebra which exists constructively. Proposition 8.24 The set Reg A over A is a Boolean residuation ideal of P ( A∗ ). Proof Let L be a regular language over A. We have seen, in Theorem 8.22, that L is recognized by a finite monoid quotient h : A∗  M. That is, L = h−1 (u) where u := h[L] ⊆ M is the direct image of L under h. It follows that the complementary language L c = A∗ – L is also recognized by h, via the set uc = M – u. Thus, the set of regular languages is closed under complementation. The set Reg A clearly contains A∗ . To see that it is closed under intersection, notice that, if L is recognized by h : A∗  M and K is recognized by g : A∗  N via v ⊆ N, then L ∩ K is recognized by the product map h, g : A∗ → M × N,

w → (h(w), g(w))

via the subset u × v of M × N. It remains to show that Reg A is a residuation ideal. For any L ∈ Reg A, the residuation ideal L(L) is finite by Theorem 8.22. Thus, for any language K ∈ P ( A∗ ), the languages K\L and L/K are in the finite residuation ideal L(L), so that the residuation ideals L(K\L) and L(L/K ) are contained in L(L) and thus also finite. By Theorem 8.22, K\L and L/K are regular, as required.  Notice that, even though we took h and g in the proof of closure under binary intersection in Proposition 8.24 to be surjective, the monoid homomorphism h, g is not necessarily surjective. Of course, we could have gotten a surjective monoid morphism by looking at the surjective co-restriction of the product map. This is one of the reasons why it is more natural to relax the notion of recognition to monoid

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

290

8 Automata Theory

morphisms rather than insisting on surjections. Also note that, by taking products of homomorphisms, as in the proof of Proposition 8.24, we obtain the following corollary. Corollary 8.25 If L 1, . . . , L n are regular languages over A, then there exists a finite monoid quotient h : A∗  M that recognizes all the languages in the Boolean subalgebra generated by L 1, . . . , L n . In the next section, we will apply topological duality to the Boolean residuation algebra (Reg A, /, \). Exercises for Section 8.1 Exercise 8.1.1 Let A be a finite alphabet and u ∈ A∗ . Show that the singleton {u} is a regular language by exhibiting a finite state automaton which recognizes this singleton language. Conclude that the Boolean algebra of all finite and co-finite subsets of A∗ is contained in Reg A. Exercise 8.1.2 Let A be a finite alphabet. For w ∈ A∗ , write c(w) for the set of letters that occur in w. Consider the language L = {w ∈ A∗ | c(w)  A}. (a)

(b) (c) (d)

Give a finite automaton with | A| states that recognizes L. Hint. In each state, verify that a particular letter does not appear in the word; make all states initial. Show that c : A∗ → P ( A) is a monoid homomorphism, where the monoid operation on P ( A) is ∪, with neutral element ∅. Prove that c is, up to isomorphism, the syntactic homomorphism of L. Conclude that L can not be recognized by a monoid that has fewer than 2 | A| elements.

Note. The syntactic monoid of a regular language L is in general isomorphic to the monoid of transitions of the so-called minimal automaton of L, see, for example, Pin (2021a, Chapter 1, Section 4.3). Also, see Exercise 8.1.10 below. Exercise 8.1.3 Let A be a one-element alphabet. (a)

Show that A∗ is isomorphic to the monoid (N, +).

In the rest of this exercise, we will identify the monoid A∗ with (N, +). (b) Show that the kernels of surjective monoid homomorphisms (N, +)  M, with M finite, are exactly the monoid congruences ϑ N,q , where, for N, q ∈ N with q ≥ 1, we write ϑ N,q for the monoid congruence on N generated by the single pair (N, N + q). Also, show that a pair (m, n) ∈ N2 is in ϑ N,q if, and only if, m = n or both m and n are ≥ N and m − n is divisible by q.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.1 The Syntactic Monoid as a Dual Space

291

(c) Conclude that if M is a finite monoid that is generated by a single element, then M is isomorphic to N/ϑ N, q for some N, q ∈ N. Note. An important fact in the theory of finite monoids, closely related to the preceding two items, is that every finite semigroup generated by a single element contains a unique idempotent; see the proof of Proposition 8.39, which contains hints for item (c). (d) Show that the collection of regular languages in N is generated as a Boolean algebra by the singleton languages and the languages of the form qN + r := {qn + r | n ∈ N}

for q, r ∈ N .

(e) Show that the language {2n | n ∈ N} is not regular. (f) Show that Reg A is not complete. Hint. Use Exercise 1.2.6. Exercise 8.1.4 Let A be a finite set and Z2 be the monoid of integers modulo 2 under addition. Consider the finite state automaton A = (Z2, A, δ, {0}, {0}), where δ = {(0, a, 1), (1, a, 0) | a ∈ A}. Show that A recognizes the language consisting of all words in A∗ of even length. Exercise 8.1.5 Prove that the identity element of a monoid is unique. Give an example of a monoid M with a subsemigroup N that is not a submonoid of M, but such that N is a monoid in its own right. Exercise 8.1.6 Let M be a monoid. Suppose that J is a collection of residuation  ideals of P (M). Prove that J is again a residuation ideal. Prove that if, moreover, the residuation ideals in J are all Boolean and/or complete, then the same is true  for J . Exercise 8.1.7 This exercise fills in the details of Proposition 8.15. Let M be a monoid, S a set, and let λ : M × S → S and ρ : S × M → S be functions. For any m ∈ M, write λ m (s) := λ(m, s) and ρm (s) := ρ(s, m). Let r : P (S) × M → P (S) and  : M ×P (S) → P (S) be the functions defined, for any m ∈ M and u ∈ P (S), by r (u, m) := λ −1 m (u), (a) (b) (c)

(m, u) := ρ−1 m (u).

Prove that r is a right action of M on P (S) if, and only if, λ is a left action of M on S. Conclude that also  is a left action if, and only if, ρ is a right action. Prove that (λ, ρ) is a biaction of M on S if, and only if, (, r) is a biaction of M on P (S). Prove that, if (λ , ρ ) is a biaction of M on S  with dual biaction ( , r  ) on P (S  ), then h : S → S  is a biaction morphism if, and only if, h−1 : P (S  ) → P (S) is a biaction morphism.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

292

8 Automata Theory

Exercise 8.1.8 Show that the inclusion (N, +) → (Z, +) is an epimorphism in the category of monoids and homomorphisms, although this is clearly not a surjective function. Exercise 8.1.9 Prove that the function δ defined before Lemma 8.5 is a homomorphism of monoids. Exercise 8.1.10 Let M be a monoid and let s ∈ P (M). (a)

Show that the dual of the complete Boolean residuation ideal generated by s is given by hs : M  M/ ≡s, where m ≡s m  if, and only if, ∀x, y ∈ M

(b)

(xmy ∈ s ⇐⇒ xm  y ∈ s).

Further, show that this is the least recognizer of s in the sense that if h : M → N is a homomorphism of monoids, then there is a unique map h˜ : im(h)  M/ ≡s such that h˜ ◦ h  = hs , where h  : M  im(h) is the surjective corestriction of h to its image.

Exercise 8.1.11 A consequence of (the proof of) Theorem 8.22, which gets a bit lost between the lines because the proof is construed as a cycle around five conditions, is the following fact which we ask you to prove directly. Let h : M → N be a homomorphism between monoids and let u ⊆ N and s ⊆ M. (a)

Show directly from the definitions that s\h−1 (u) = h−1 (h[s]\u).

(b)

Conclude that, if h is surjective, then for any p ⊆ N we have h−1 (p)\h−1 (u) = h−1 (p\u).

(c)

Show that, if N is finite, then there exist m1, . . . , mk ∈ s with s\h−1 (u) =

k

mi−1 h−1 (u).

i=1

8.2 Regular Languages and Free Profinite Monoids In Section 8.1, we considered regular languages one at a time, and we saw that the dual of the residuation ideal generated by such a language is finite – essentially by virtue of the simple fact that any finite state machine can only be equipped with finitely many choices of pairs (I, F) of initial and final states. We placed

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

293

this result within discrete duality, and, as a consequence, we were able to see that regular languages over an alphabet A are precisely those subsets of P ( A∗ ) that are recognized by finite monoids, and that the collection of regular languages is a Boolean residuation ideal of P ( A∗ ). In the current section, we want to consider the set of all regular languages together as a whole. Since the Boolean residuation ideal of regular languages over A is not complete (see Exercise 8.1.3), we have to switch from the discrete duality to Stone’s topological duality. We will be rewarded by seeing that its dual is a very natural object from topological algebra, namely the free profinite monoid over A. Next, in Section 8.3, we apply the subalgebra–quotient-space duality to the pair (regular languages over A, the free profinite monoid over A) focusing in on subalgebras of regular languages that are residuation ideals. This subject culminates in the pairing of so-called pseudovarieties of regular languages and relatively free profinite monoids that is known in automata theory via the combination of Eilenberg’s theorem and Reiterman’s theorem. We will not give the general theory, but we will illustrate with the example of so-called piecewise testable languages. In particular, we will use the duality to give a proof of Simon’s theorem, which gives a decidable characterization of the piecewise testable languages among the regular ones via profinite monoid equations. Finally, in Section 8.4, we again apply the subalgebra–quotient-space duality to the pair (regular languages over A, the free profinite monoid over A) but this time we focus on subalgebras of regular languages that are closed under concatenation. This points in the direction of categorical logic and hyperdoctrines rather than automata theory and is beyond the scope of this book. We just give a few elementary observations in the short final section.

Duality for the Boolean Residuation Algebra of Regular Languages Our main focus in this subsection will be to prove the following theorem; the terms “Boolean residuation algebra” and “free profinite monoid” will be defined below. Theorem 8.26 Let A be a finite set. Then, Reg A is a Boolean residuation ideal of P ( A∗ ) and the dual space of the Boolean residuation algebra (Reg A, \, /) is the free profinite monoid over A. In order to regard Reg A as an algebraic structure in its own right, without necessarily having to refer to its representation inside the algebra P ( A∗ ), we now introduce the following notion of Boolean residuation algebra, a Boolean algebra

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

294

8 Automata Theory

equipped with two implication-type operators, that are linked to each other via a Galois property. 4 Definition 8.27 A Boolean residuation algebra is a tuple (B, \, /), where B is a Boolean algebra and \, / : B 2 → B are binary operations with the following properties: (a) (b) (c)

the operation \ preserves finite meets in the second coordinate, that is, a\ =  and a\(b1 ∧ b2 ) = (a\b1 ) ∧ (a\b2 ) for all a, b1, b2 ∈ B; the operation / preserves finite meets in the first coordinate, that is, /b =  and (a1 ∧ a2 )/b = (a1 /b) ∧ (a2 /b) for all a1, a2, b ∈ B; and the two operations \ and / are linked by the following Galois property: for all a, b, c ∈ B, b ≤ a\c if, and only if, a ≤ c/b.

Note that, under the assumption of (c), conditions (a) and (b) in Definition 8.27 are equivalent. The Galois property (c) also implies that \ and / are implicationtype operators, in the sense of Definition 4.58, that is, the following equations automatically hold: • ⊥\b =  and (a1 ∨ a2 )\b = (a1 \b) ∧ (a2 \b) for all a1, a2, b ∈ B and • a/⊥ =  and a/(b1 ∨ b2 ) = (a/b1 ) ∧ (a/b2 ) for all a, b1, b2 ∈ B. Now, for the following corollary to Proposition 8.24 of the previous section, simply note that, since Reg A is closed under \ and / with arbitrary denominators from P ( A∗ ), it is in particular a residuation algebra in its own right. Corollary 8.28 Let A be a finite set. Then, Reg A, the set of all regular languages over A, is a Boolean residuation algebra. Remark 8.29 Recall that P ( A∗ ) is a residuated Boolean algebra, that is, a Boolean algebra equipped with a monoid operation which preserves join that is residuated (see Exercise 4.5.3). In fact, Reg A is not only a residuation algebra, but even a residuated Boolean subalgebra of P ( A∗ ). That is, in addition to being closed under the residuation operations, it is also closed under concatenation product of languages. However, Reg A being closed under concatenation is a different phenomenon than what we are looking at here. In fact, for any monoid M, the set of all the subsets of M recognized by finite monoids is a residuation ideal in P (M) and therefore, in particular, closed under the residual operations, but this Boolean subalgebra is in general not closed under the concatenation product of P (M) (see Exercise 8.2.11). We will study closure under concatenation product in Section 8.4. 4

In this book, for simplicity we only consider the case where the operators are binary, although a more general version exists in the literature, see Gehrke (2016).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

295

Now, we are ready to consider the dual space of the Boolean residuation algebra Reg A. We will prove that this dual space is equipped with a monoid operation such that it is, up to isomorphism, the free profinite monoid on A. We now give the necessary definitions. Definition 8.30 A topological monoid is a tuple (M, τ, ·), where (M, τ) is a topological space, (M, ·) is a monoid, and the function · : M 2 → M is continuous. A discrete monoid is a topological monoid whose topology is discrete. A topological monoid is profinite if it is a projective limit of finite discrete monoids in the category of topological monoids. An iso-homeomorphism between topological monoids is a function that is both a homeomorphism of the underlying topological spaces and an isomorphism of the underlying monoids. Unless mentioned otherwise, a finite monoid is always equipped with the discrete topology. Let us first give a more concrete characterization of profinite monoids, that is often convenient to work with. When M is a topological monoid, we will denote by ClpCon(M) the set of clopen congruences on M, that is, congruences ϑ such that ϑ is clopen as a subset of M × M. Recall from Exercise 4.2.10 that for a congruence ϑ to be clopen, it is equivalent to say that M/ϑ is finite and M  M/ϑ is continuous for the discrete topology on M/ϑ. Note further that, for a congruence ϑ such that M/ϑ is finite, the continuity condition is equivalent to requiring that each equivalence class [m]ϑ is clopen. Proposition 8.31 A topological monoid M is profinite if, and only if, the space underlying M is compact, and, for every x, y ∈ M, if x  y, then there exists a continuous homomorphism f : M → F, with F a finite discrete monoid, such that f (x)  f (y).  If M is profinite, then M is isomorphic to the closed submonoid of ϑ ∈ClpCon(M ) M/ϑ consisting of the tuples (x ϑ )ϑ ∈ClpCon(M ) such that, whenever ϑ ⊆ ϑ  and m ∈ M is such that [m]ϑ = x ϑ , then [m]ϑ  = x ϑ  . Proof See Exercise 8.2.8.



Remark 8.32 Recall that Boolean topological spaces are the same thing as profinite sets; see Example 5.35 in Chapter 5. In particular, it is easy to see from Proposition 8.31 that the topology of a profinite monoid must be Boolean, since it is a closed submonoid of a product of finite spaces. It is less immediate that the converse holds true, in the following sense: If · is a continuous associative multiplication on a Boolean space M with neutral element 1, then (M, ·, 1) is a profinite monoid. Toward showing the condition in the first part of Proposition 8.31, if x  y, there is, by zero-dimensionality of the space, a clopen set separating x from y, but it remains to show that this clopen set can be realized as a union of equivalence classes

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

296

8 Automata Theory

for some ϑ ∈ ClpCon(M). This is true in the case of monoids, where it is known as “Hunter’s lemma,” but it does not generally hold true for other algebraic structures based on Boolean spaces; see Almeida et al. (2023) for a recent analysis of the question of what makes a topological algebra based on a Boolean space profinite. The characterization of Proposition 8.31 in particular allows us to define an element of a profinite monoid by giving its value modulo ϑ, for each ϑ ∈ ClpCon(M). This idea will be crucial below for showing that the dual space of Reg A is a free profinite monoid over A. Definition 8.33 Let A be a finite set. A free profinite monoid over A is a profinite monoid X together with a function η : A → X, such that, for any finite monoid M and function f : A → M, there exists a unique continuous homomorphism f: X → M such that f◦ η = f , as in the following diagram: η

A

X  f

f

M. In Definition 8.33, the continuity of fis with respect to the discrete topology on M and means exactly that f−1 (m) is clopen for every m ∈ M. If (X, η) and (X , η  ) are free profinite monoids over A, then there is a unique iso-homeomorphism ϕ between X and X  such that ϕ ◦ η = η , see Proposition 8.31 and Exercise 8.2.9. The fact that free profinite monoids exist can be shown with a construction in the same vein as the one for profinite ordered sets in Remark 5.36, by taking the limit of a diagram of (discrete) finite monoids in the category of topological monoids. However, our proof that the free profinite monoid exists here will be by exhibiting it as the dual space of the Boolean residuation algebra (Reg A, \, /). For the rest of this subsection, let A be a finite set and let X denote the dual space of the Boolean algebra Reg A. Toward proving that X is a free profinite monoid over A, note first that, since Reg A is a Boolean subalgebra of P ( A∗ ), its dual space X is a topological quotient of the dual space of P ( A∗ ), which is the Stone–Čech compactification β A∗ of the set A∗ . In particular, we have a natural map A∗ → X given by sending u ∈ A∗ to the point of X corresponding to the ultrafilter consisting of all regular languages that contain the word u. Since the singleton {u} is regular, this map is injective and the image consists entirely of isolated points (see Exercises 8.1.1 and 8.2.2). In other words, the map A∗ → X embeds the discrete space A∗ in X. We will henceforth identify A∗ with its image in X and consider A∗ as contained in X. Remark 8.34 In what follows, we will use the following consequence of Stone duality, which is the “topological half” of the statement that X is the free profinite

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

297

monoid over A (see Exercise 8.2.7 for a proof). Let h : A∗  M be a surjective homomorphism to a finite monoid M. Viewing the finite set M as a discrete topological space, the function h has a unique continuous extension h¯ : X  M, ¯ = h(w) for all w ∈ A∗ . Further that is, h¯ −1 (m) is clopen for every m ∈ M and h(w) note that, if  ∈ Reg A is recognized by h via P ⊆ M, that is, if  = h−1 (P), then the −1 (P) = h ¯ −1 (P) (see Exercise 8.2.7(d)). corresponding subset of X is equal to h To properly formulate and prove the “monoid half” of the statement that X is the free profinite monoid over A, we first need to construct a monoid operation on X. This monoid operation comes from the dual of the residuation operations \ and / of the Boolean residuation algebra Reg A. Recall that in Definition 4.65 we defined a ternary relation dual to any implication-type operator between distributive lattices. Instantiating (4.22) in that definition for the specific implication-type operator \ : Reg A × Reg A → Reg A, we obtain the ternary relation R\ on X defined by  then z ∈  k. R\ (x, y, z) ⇐⇒ for all , k ∈ Reg A, if x ∈ and y ∈ \k,

(8.2)

Lemma 8.35 Let x, y, z ∈ X. Then, R\ (x, y, z) if, and only if, for any finite monoid ¯ ¯ h(y). ¯ = h(x) quotient h : A∗  M, we have h(z) Proof For the left-to-right direction, suppose R\ (x, y, z) and let h : A∗  M be a finite monoid quotient with unique continuous extension h¯ : X → M. Write ¯ ¯ ¯ p := h(x) and q := h(y); we need to show that h(z) = pq. Consider the regular −1 −1 languages  := h (p) and k := h (pq). Using Remark 8.34, since  = h¯ −1 (p),  Also, we have x ∈ . \k = h−1 (p−1 {pq}) by Exercise 8.1.11, using the fact that h is a monoid homomorphism. Since obvi using Remark 8.34 again. Now, ously q ∈ p−1 {pq}, we get y ∈ h¯ −1 (p−1 {pq}) = \k, k = h¯ −1 (pq), as required. because R\ (x, y, z) by assumption, we must have z ∈  ¯ ¯ h(y) ¯ For the right-to-left direction, suppose h(z) = h(x) for all finite monoid ∗ quotients h : A  M. To show R\ (x, y, z), suppose that , k ∈ Reg A are such  we need to prove z ∈  that x ∈  and y ∈ \k; k. By Corollary 8.25, pick a finite ∗ monoid quotient h : A  M which recognizes both languages  and k, that is, pick P, Q ⊆ M such that  = h−1 (P) and k = h−1 (Q). By Remark 8.34, we have  we have h(x) ¯  = h¯ −1 (P) and  k = h¯ −1 (Q). Since x ∈ , ∈ P. Also, using Exercise 8.1.11, \k = h−1 (P)\h−1 (Q) = h−1 (p\Q)  = h¯ −1 (P\Q) and thus h(y) ¯ ∈ P\Q. It now follows from the definition so that \k ¯ ¯ ¯ ¯ h(y) ¯ of P\Q that h(x) h(y) ∈ Q. Now, since h(z) = h(x) by hypothesis, we get −1  ¯  z ∈ h (Q) = k, as required.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

298

8 Automata Theory

Lemma 8.36 Let x, y ∈ X. The set of regular languages μ x,y := { ∈ Reg A | there is a finite monoid quotient h : A∗  M which ¯ h(y)) ¯ ⊆ } recognizes  and such that h−1 ( h(x) is an ultrafilter of Reg A. Moreover, the point z ∈ X for which Fz = μ x,y is the unique point in the set R\ [x, y, _]. Proof Note first that, for any  ∈ Reg A, if h : A∗  M is a finite monoid quotient which recognizes , then, for any m ∈ M, exactly one of the following two properties holds: h−1 (m) ⊆ 

or

h−1 (m) ⊆ A∗ –  .

Indeed, at most one of the two can hold because h−1 (m) is non-empty, since h is surjective. Also, since h recognizes , we have  = h−1 (u) for some u ⊆ M, so that at least one of the two must hold: the first in case m ∈ u and the second in case m  u. It follows that, for every regular language , since there exists some finite monoid quotient h : A∗  M recognizing , we have that either  or A∗ –  is in μ x,y , ¯ h(y). ¯ applying the above argument to m := h(x) It is also clear, by the surjectivity of h, that ∅  μ x,y . It remains to prove that μ x,y is a filter. To this end, let  1,  2 be regular languages. If  1 and  2 are both in μ x,y , then, for i = 1, 2, pick finite monoid quotients hi : A∗  Mi , recognizing  i and such that hi−1 (hi (x)hi (y)) ⊆  i . Define the homomorphism h : A∗  M as the co-restriction of the homomorphism h1 × h2 : A∗ → M1 × M2 to its image. Then, h is a finite monoid quotient of A∗ that recognizes  1 ∩  2 , as we saw in the proof of Proposition 8.24. Moreover, for i = 1, 2, we have h−1 (h(x)h(y)) ⊆ hi−1 (hi (x)hi (y)) . Combining this with the assumption that hi−1 (hi (x)hi (y)) ⊆  i , we see that h−1 (h(x)h(y)) ⊆  1 ∩  2 , so that  1 ∩  2 ∈ μ x,y . To see that μ x,y is an up-set, suppose that  1 ∈ μ x,y and that  1 ⊆  2 . Pick a finite monoid quotient h1 : A∗  M1 recognizing  1 such that h−1 (h1 (x)h1 (y)) ⊆  1 and pick some finite monoid quotient h2 : A∗  M2 recognizing  2 . Defining h : A∗  M in the same way as before, this morphism still recognizes  2 and, since h−1 (h(x)h(y)) intersects non-trivially with  1 , it also intersects non-trivially with  2 . For the moreover statement, denote by z the unique point of X such that Fz = μ x,y . We use Lemma 8.35 to show that R\ (x, y, _) = {z}. To see that R\ (x, y, z), let ¯ h(y)) ¯ is h : A∗  M be a finite monoid quotient. Then, the language  := h−1 ( h(x) −1  ¯ ¯ ¯ regular and, clearly,  ∈ μ x,y = Fz , which means that z ∈  = h ( h(x) h(y)). We conclude that R\ (x, y, z) by Lemma 8.35. Conversely, if R\ (x, y, z  ) for any point

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

299

z  ∈ X, then let  ∈ μ x,y be arbitrary. Pick a finite monoid quotient h : A∗  M ¯ h(y)) ¯ ¯ h(y)). ¯ ⊆ . Write   := h−1 ( h(x) By Lemma recognizing  with h−1 ( h(x)      ¯ ¯ ¯ 8.35, h(z ) = h(x) h(y), so z ∈  ⊆ . Thus, μ x,y ⊆ Fz , and the two ultrafilters  must be equal, so z  = z. We summarize Lemma 8.35 and Lemma 8.36 in the following corollary. Corollary 8.37 The ternary relation R\ dual to \ is a total binary operation : X × X → X and the ultrafilter corresponding to x  y is μ x,y . Moreover, for any x, y ∈ X, x  y is the unique element of X such that, for every finite monoid ¯  y) = h(x) ¯ h(y). ¯ quotient h : A∗  M, we have h(x Of course, we also have the ternary relation R/ dual to the operator /, which is of “reverse” implication type, in the sense that it is monotone in the first and antitone in the second coordinate and satisfies (4.19) and (4.20) in Definition 4.58 with the first and second coordinate swapped. Concretely, we may define, analogously to (8.2),  then z ∈  k. R/ (x, y, z) ⇐⇒ for all , k ∈ Reg A, if x ∈ and y ∈ k/,

(8.3)

However, this ternary relation R/ does not add any additional structure to the dual space, as it is closely related to R\ above. Indeed, the Galois property in the definition of the Boolean residuation algebra Reg A (see Definition 8.27) expresses the fact that the relations R/ and R\ are related via R\ (x, y, z) ⇐⇒ R/ (y, x, z) for any x, y, z ∈ X, so that R/ essentially contains the same information as R\ . This may be proved in a similar way to the case of relations dual to an adjunction considered in Exercise 4.3.5 (see Exercise 8.2.3). Finally, we show that the operation  makes X into a topological monoid and that it satisfies the universal property of the free profinite monoid over A. Recall that we view A∗ as a discrete subspace of X and we denote the empty word by . Lemma 8.38 The triple (X, ,  ) is a free profinite monoid over A. Proof First we show that (X, ) is a monoid with neutral element the empty word  ∈ A∗ ⊆ X. Let x, y, z ∈ X. In order to prove (x  y)  z = x  (y  z), we show that the points have the same clopen neighborhoods. Let  ∈ Reg A be arbitrary and  we will show that x  (y  z) ∈ .  Pick h : A∗  M suppose that (x  y)  z ∈ ; a finite monoid quotient which recognizes  via P, so that  = h¯ −1 (P). Repeatedly using the defining property of  in Corollary 8.37, we have ¯ ¯ h(y)) ¯ ¯ ¯  (y  z)) = h(x)( ¯ ¯ h(z)). ¯ h((x  y)  z) = ( h(x) h(z) and h(x h(y)

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

300

8 Automata Theory

Now, since M is a monoid, these two elements of M equal, and thus in particular  Similar reasoning shows that  is a neutral we also have x  (y  z) ∈ h¯ −1 (P) = . element for  (see Exercise 8.2.4). Next, we need to show that the function  is continuous. Suppose x, y ∈ X,  Again, let h : A∗  M be a finite quotient which  ∈ Reg A and x  y ∈ . ¯ ¯ and  2 = h−1 ( h(y)). recognizes  via P, so that  l = h¯ −1 (P). Let  1 = h−1 ( h(x))     Then,  1,  2 ∈ Reg A and, for any x ∈  1 and y ∈  2 , we have ¯   y  ) = h(x ¯  ) h(y ¯  ) = h(x) ¯ h(y) ¯ ¯  y) ∈ P. h(x = h(x  as required. That is, x   y  ∈ , We now show that it follows from the characterization in Proposition 8.31 that the topological monoid X is profinite. Indeed, if x and y are distinct elements of X, then  Pick h : A∗ → M a homomorphism there exists  ∈ Reg A such that x ∈ and y  . to a finite monoid that recognizes , say  = h−1 (P) for some P ⊆ M. By Corollary 8.37, the unique continuous extension h¯ : A∗ → M is a homomorphism, and since ¯ ¯ ¯ ¯ h(x) ∈ P and h(x)  P, we must have that h(x)  h(y). Finally, we show that (X, ,  ) has the universal property of a free profinite monoid over A with respect to the function η which sends each letter a ∈ A to the one-letter word a in X. To this end, let M be a finite monoid and let f : A → M be any function. Then, since A∗ is the free monoid over A, we obtain a unique homomorphic extension f ∗ : A∗ → M. Since im( f ∗ ) is closed in the discrete space M, any continuous extension of f ∗ will map onto im( f ∗ ), so it suffices to consider the surjective co-restriction h : A∗  im( f ∗ ) of f ∗ to its image. As we have been using extensively throughout, see Remark 8.34, h extends uniquely to a continuous map h¯ : X  im( f ∗ ), and, by Corollary 8.37, h¯ is also a monoid morphism with respect to . In a diagram, A

A∗

X h

f

f∗

h¯

im( f ∗ )

M. Each of the inner triangles of the diagram commutes and thus so does the outer triangular diagram. Finally, the outer vertical arrow X → M is the unique such continuous homomorphism, since A generates A∗ , A∗ is dense in X, and h¯ is both a monoid homomorphism and continuous (see Exercise 8.2.5).  This completes the proof of Theorem 8.26. We chose to give a very concrete proof of this theorem as this is maybe more hands-on and tangible for the novice and because such a proof is not readily available in the literature. Usually, in research

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

301

papers, a more conceptual proof is given: Reg A is clearly the filtered colimit of its finite residuation ideals, so, by duality, its dual space is the cofiltered, or projective, limit of the duals of these finite residuation ideals, which, by Theorem 8.22, are the finite monoid quotients of A∗ . It remains to check that the projective limit of the diagram of finite monoid quotients of A∗ , taken in the category of topological monoids, is indeed a free profinite monoid. This yields a quicker, simpler, but maybe somewhat magical proof of Theorem 8.26. See, for example, Gehrke (2016, Theorem 4.4) for further details of this alternate proof. Now that we know that a free profinite monoid over A exists and is unique up to iso-homeomorphism, we can speak of the free profinite monoid over A, and we will denote it by  A∗ . We now give one more instance of a definition of an operation on a profinite monoid “via the finite quotients,” which we will use in Section 8.3. An element e of a monoid is called idempotent if e · e = e. Proposition 8.39 Let M be a profinite monoid. For any x ∈ M, there is a unique idempotent element in the closure of the set {x n | n ≥ 1}. Proof First note that, for a finite monoid M, there is always a unique idempotent element in the set {x n | n ≥ 1}, for every x ∈ M. Indeed, by the pigeon-hole principle, pick n, p ≥ 1 such that x n = x n+p . Note that it follows that, for any q ≥ 1, we have x n+qp = x n . Let r be the remainder of the division of −n by p, that is, choose the unique 0 ≤ r < p such that n + r is a multiple of p. Then, x n+r is idempotent, as shown by the computation: x n+r x n+r = x n+r+qp = x n+qp x r = x n x r = x n+r . If, for any m ≥ 1, x m is also idempotent, then x m = (x m ) n+r = (x n+r ) m = x n+r , so this idempotent is unique. This establishes the proposition for the special case of finite (discrete) monoids. Now, if M is a profinite monoid and x ∈ M, let us write P := {x n | n ≥ 1}. If y is an idempotent element in the closure of P, then for any continuous homomorphism f : M → F, with F a finite discrete monoid, f (y) must be idempotent, and since y is in the closure of P, it must be in the closed set f −1 ( f [P]), so that f (y) is a power of f (x). By Proposition 8.31 and the fact that f (x) has a unique idempotent power for every f , there can be at most one such element y. Moreover, such an element exists, because if ϑ ⊆ ϑ  and [x]ϑm is idempotent, then [x]ϑm is idempotent, as well. So, using the second part of Proposition 8.31 we have the element x ω of M defined by the condition that, for each ϑ ∈ ClpCon(M), [x ω ]ϑ is the unique idempotent  power of [x]ϑ in M/ϑ. Definition 8.40 Let M be a profinite monoid. For x ∈ M, we denote by x ω the unique idempotent element in the closure of {x n | n ≥ 1}.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

302

8 Automata Theory

Similarly, there exists, for any x ∈ M, an element x ω−1 in the closure of {x n | n ≥ 1} which is uniquely determined by the condition that x ω−1 · x = x · x ω−1 = x ω (see Exercise 8.2.10). A General Duality for Boolean Residuation Algebras Above, we have seen that Stone duality on objects restricts to a correspondence between finitely generated free profinite monoids and the Boolean residuation algebras of the form Reg A, for A a finite alphabet. In order to show that this is a more general phenomenon, we will show here that (a) (b)

all binary topological algebras whose underlying space is Boolean are dual spaces of Boolean residuation algebras and the residuation algebras dual to binary topological algebras are precisely those which preserve joins at primes.

Item (b) is particularly interesting as it makes a link to domain theory in logical form as treated in Sections 6.4 and 7.4, where the concept of preserving joins at primes is also central, see Corollary 6.34 and Theorem 7.69. By a binary topological algebra, we mean a pair (X, f ), where X is a topological space and f : X 2 → X is a continuous binary operation on X, with no equational axioms assumed on f . A set equipped with a binary operation satisfying no axioms has also been called magma in the algebraic and categorical literature; from that perspective, we here consider magmas internal to the category Top, and in particular in the full subcategory BoolSp. The fact that we restrict ourselves to binary operations is just for simplicity of notation and a more general theorem for arbitrary arity can be proved, see Gehrke (2016, Section 3). We will use the term “binary topological algebra on a Boolean space” for one whose underlying space is Boolean; these structures could also be called “Boolean-topological magmas.” Recall from Section 4.5 that implication-type operators on a lattice are dual to compatible ternary relations on its dual space. Thus, to show (a), we start by showing that a binary operation on a Boolean space is continuous if, and only if, its graph is a compatible ternary relation in the sense of Definition 4.63. By the graph of a binary operation f : X 2 → X, we here mean the ternary relation {(x, y, f (x, y)) | x, y ∈ X }. Since we are in the special case of Boolean spaces, we will here specialize the definition of compatibility for a ternary relation to that setting. Note that a relation R ⊆ X 3 is then compatible of implication type (Definition 4.63) if for any clopen subsets U, V of X, R[U, _, V ] is clopen and for any y ∈ X, the set R[_, y, _] is closed.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

303

Proposition 8.41 Let X be a Boolean space, f : X 2 → X a binary operation on X, and R ⊆ X 3 its graph. Then, f is continuous if, and only if, its graph is a compatible relation on X. Proof Throughout this proof, for any y ∈ X, we denote by i y : X → X 2 the injective function x → (x, y), which is continuous by Exercise 2.1.13, and we write f y := f ◦ i y , that is, f y : X → X is the function defined by f y (x) = f (x, y) for every x ∈ X. Suppose first f is continuous. Then, f y is also continuous, and for any y ∈ X, the binary relation R[_, y, _] is the graph of the function f y : X → X. This graph is closed because X is Hausdorff (see Exercise 2.1.12). Let U, V ⊆ X be clopen. First notice that R[U, _, V ] = π2 [(U × X ) ∩ f −1 (V )], where we denote by π2 : X 2 → X the projection on the second coordinate. Further, since X is compact, the projection π2 is a closed mapping by Proposition 2.5, and thus R[U, _, V ] is closed. Finally, by Exercise 2.1.13, π2 is always an open mapping, and thus R[U, _, V ] is clopen, as required. For the converse, suppose the graph R is compatible. We show first that f y is continuous for every y ∈ X. For any S ⊆ X, we have f y−1 (S) = π1 [R[_, y, _] ∩ (X × S)], that is, the inverse image of S under f y is the same as the direct image under π1 of the set R[_, y, _] ∩ (X × S). Again, since π1 is a closed mapping and R[_, y, _] is closed, this shows that f y−1 (S) is closed whenever S is closed. Now, let V ⊆ X be clopen; we show that f −1 (V ) is open. Let (x, y) ∈ f −1 (V ) be arbitrary. Then, x ∈ f y−1 (V ), which is open, so pick a clopen U ⊆ X satisfying x ∈ U ⊆ f y−1 (V ). Consider the set U\V := R[U, _, V c ]c = {w ∈ X | for all u ∈ U, f (u, w) ∈ V }, which is clopen because R is compatible. Note that, since U ⊆ f y−1 (V ), we have y ∈ U\V and, clearly, U × (U\V ) ⊆ f −1 (V ). Thus, U × (U\V ) is an open neighborhood around (x, y) contained in f −1 (V ) and we conclude that f −1 (V ) is open, as required.  Notice that by symmetry of assumptions, exactly the same proof with the role of the first two coordinates of the relation R switched would work just as well. Thus, if f is a continuous operation on X, then its dual Boolean algebra is a residuation algebra. We get the following corollary. Corollary 8.42 Every binary topological algebra on a Boolean space is the dual space of a Boolean residuation algebra.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

304

8 Automata Theory

For later use in this section, we now make one direction of the object duality of Corollary 8.42 more explicit. Notation Throughout the rest of this section, in order to simplify notation, when B is the Boolean algebra of clopens of a Boolean space X, we will tacitly identify ultrafilters of B with points of X; that is, if U ∈ B and x ∈ X then U belongs to the ultrafilter corresponding to x if, and only if, x ∈ U. Let (X, f ) be a binary topological algebra on a Boolean space, let R be the graph of f , and let B be the Boolean algebra of clopens dual to X. The algebra dual to (X, R) is then the Boolean residuation algebra (B, \, /) where, if U, V ∈ B, then U\V = R[U, _, V c ]c = {y ∈ X | for all u ∈ U, f (u, y) ∈ V }. Here, we have simply applied the definition of implication-type operator associated with a ternary relation (4.21) in the specific case where the relation is the graph of a continuous binary function. The definition of the other residual / is similar but with the role of the first two coordinates switched. When U = {x}, we will use x −1V as a notation for {x}\V . We now proceed to characterize exactly which Boolean residuation algebras can occur as the algebra dual to a binary topological algebra on a Boolean space. This is where the notion of preserving joins at primes reappears. Let (B, \, /) be a Boolean residuation algebra, as defined in Definition 8.27. Also, recall from Definition 6.32 what it means for an implication-type operator to preserve joins at primes. That definition was given for an implication-type operator → : L×M → K, which is antitone in the first and monotone in the second coordinate, and thus directly specializes to the residual \ on a Boolean residuation algebra: The operation \ preserves joins at primes if, for every ultrafilter F of B, a ∈ F, and finite subset G of B, there exists a  ∈ F such that   (a \g). a\ G ≤ g ∈G

The other residual / is of “reverse” implication type. We say that the operation / preserves joins at primes if for every ultrafilter F of B, a ∈ F, and finite subset G of B, there exists a  ∈ F such that  

(g/a  ). G /a ≤ g ∈G

Definition 8.43 Let (B, \, /) be a Boolean residuation algebra and let X be its dual. We say that the algebra B preserves joins at primes provided that both operations \ and / preserve joins at primes.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

305

It will in fact follow from the proof of Theorem 8.45 that the one residual / preserves joins at primes if, and only if, the other residual \ preserves joins at primes, so it would suffice in Definition 8.43 to assume only one of the two. Before proving Theorem 8.45, we isolate one lemma that contains an important step. Lemma 8.44 Let (X, f ) be a binary topological algebra on a Boolean space X and let (B, \, /) be the dual Boolean residuation algebra of clopen subsets of X. For any x ∈ X and V ∈ B, there exists U ∈ B such that x ∈ U and x −1V = U\V . Proof Let V ∈ B and x ∈ X. We first show that the following collection of clopen subsets of X C := {K\V | K ∈ B and x ∈ K } is a cover of the set x −1V . Indeed, if y ∈ x −1V , then f (x, y) ∈ V . Reasoning as in the proof of Proposition 8.41, by continuity of the function f y : x → f (x, y), since x ∈ f y−1 (V ), there is a clopen set K around x such that K ⊆ f y−1 (V ), which means that y ∈ K\V . Now, since all sets in C are clopen, and x −1V is closed (even clopen) and hence compact, pick a finite set K ⊆ B such that x ∈ K for every K ∈ K and

 x −1V ⊆ K ∈K K\V . Define U := K . Then, clearly x ∈ U, so U\V ⊆ x −1V and, for every K ∈ K , we have that U ⊆ K, so K\V ⊆ U\V . Thus,  K\V ⊆ U\V ⊆ x −1V, x −1V ⊆ K ∈K

from which we conclude that

x −1V

= U\V , as required.



Theorem 8.45 The dual algebras of binary topological algebras on Boolean spaces are precisely the Boolean residuation algebras preserving joins at primes. Proof First, suppose that (X, f ) is a binary topological algebra and (B, \, /) is its dual Boolean residuation algebra. We need to show that (B, \, /) preserves joins at primes. We only show this for \, the argument for / is symmetric. Let F be an ultrafilter in B, which, by Stone duality, is of the form {U ∈ B | u ∈ U} for some u ∈ X. By induction, it suffices to treat just the cases where G is empty or G contains two elements; when G contains one element there is nothing to do. If G is empty, then the supremum of G in B is the empty set. Let a ∈ F and suppose that u ∈ a. We have a\∅ ⊆ {y ∈ X | f (u, y) ∈ ∅} = ∅, a

so we may take := a in the definition of preserving joins at primes. Now, suppose that G = {V1, V2 } and let a ∈ F, so that u ∈ a. Note first that a\(V1 ∪ V2 ) ⊆ u−1 (V1 ∪ V2 ) = u−1V1 ∪ u−1V2, using the fact that V → u−1V is a Boolean algebra homomorphism. By Lemma 8.44, pick U1, U2 ∈ B such that u ∈ Ui and u−1Vi = Ui \Vi for i = 1, 2. Now define

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

306

8 Automata Theory

a  := U1 ∩ U2 , which still contains u. Plugging this into the previous inclusion, we get a\(V1 ∪ V2 ) ⊆ (U1 \V1 ) ∪ (U2 \V2 ) ⊆ (a \V1 ) ∪ (a \V2 ), using for the second inclusion the fact that \ is antitone in its first coordinate. For the converse, suppose that (B, \, /) is a Boolean residuation algebra such that \ preserves joins at primes and let (X, R) be its dual. We will show that R is the graph of a binary function f : X 2 → X, that is, that R[x, y, _] is a singleton for all x, y ∈ X. Let x, y ∈ X be arbitrary. Note that it follows from the definition of R as the ternary relation dual to the implication-type operator \ (Definition 4.65) that,  for any z ∈ X, we have R(x, y, z) if, and only if, z ∈ F, where F := {V ∈ B | there exists U ∈ B such that x ∈ U and y ∈ U\V }.  We show that F is an ultrafilter, from which it follows that F is a singleton, since X is a Boolean space (also see Exercise 3.3.3). Using the fact that, for any U ∈ B, the function V → U\V preserves finite meets, one may prove that F is a filter because it is a directed union of filters (see Exercise 8.2.13). To get that F is an ultrafilter, we use the assumption that \ preserves joins at primes. First, we show that F is proper. Indeed, for any U ∈ B such that x ∈ U, the definition of preserving joins at primes applied in the case G = ∅ gives that U\∅ = ∅, so in particular y  U\∅, so that ∅ is not in F. Now, let V ∈ B be arbitrary. We show that either V or V c is in F. Applying the assumption that \ preserves joins at primes to the element  ∈ B and G = {V, V c }, since x ∈  and \(V ∪ V c ) = \ = , we can pick U ∈ B with x ∈ U and (U\V ) ∪ (U\V c ) = . Then, either y ∈ U\V or y ∈ U\V c , so one of V and V c is in F, as required. Finally, let f : X 2 → X be the function defined, for x, y ∈ X, by taking f (x, y) the unique element in R[x, y, _]. Since the relation R is compatible, it follows from Proposition 8.41 that f is continuous.  Exercises for Section 8.2 Exercise 8.2.1 Prove the remarks directly following Definition 8.27. Exercise 8.2.2 Let W be a set and let B be a Boolean subalgebra of P (W ). Denote the dual space of B by X, which is a topological quotient of βW . Consider the composition i : W → βW  X. Assume further that B contains {w} for every w ∈ W . Show that i is injective, that its image is a dense subspace of X, and that the subspace topology on i[W ] is discrete. Using Exercise 8.1.1, conclude in particular that A∗ is dense in the dual space of Reg A. Exercise 8.2.3 Let (B, \, /) be a Boolean residuation algebra with dual space X. Consider the relations R\ and R/ defined by (8.2) and (8.3), respectively. Prove that, for any x, y, z ∈ X, R\ (x, y, z) ⇐⇒ R/ (y, x, z).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.2 Regular Languages and Free Profinite Monoids

307

Exercise 8.2.4 Prove that  is a neutral element for the operation  on X, following a similar proof to the first paragraph of the proof of Lemma 8.38. Exercise 8.2.5 This exercise gives some more details to show the uniqueness of the extension constructed in the proof of Lemma 8.38. (a) (b)

Let X and Y be Hausdorff topological spaces and D a dense subspace of X.  , then h = h  . Prove that if h, h  : X → Y are continuous and h|D = hD Recall from Exercise 8.2.2 that in particular A∗ is dense in the dual space X of Reg A. Use this to conclude that, for any function f : A → M with M a finite monoid, there can be at most one continuous homomorphism X → M extending f .

Exercise 8.2.6 Let X and Y be Hausdorff topological monoids and suppose that D is a dense submonoid of X. Let f : X → Y be a continuous function such that f (u) f (v) = f (uv) for every u, v ∈ D. Prove that f is a homomorphism X → Y . Exercise 8.2.7 Let h : A∗  M be a finite monoid quotient and consider the inverse image function h−1 : P (M) → Reg A. Let X denote the dual space of Reg A. (a) (b) (c)

Explain why h−1 (P) is regular for every P ⊆ M. By applying Stone duality to the Boolean algebra homomorphism h−1 , show that h extends uniquely to a continuous function h¯ : X  M. ¯ = h(w), that is, the diagram below and Show that, for any w ∈ A∗ , h(w) commutes. A∗ X h

h¯

M. (d)

Show that if  ∈ Reg A is recognized by h via P ⊆ M, then −1 (P) = h ¯ −1 (P).  = h

Exercise 8.2.8 Let M be a topological monoid. As in Proposition 8.31, we write ClpCon(M) for the collection of clopen congruences ϑ on M. We say that the clopen congruences separate points if, for every x, y ∈ M, if x  y, then there is ϑ ∈ ClpCon(M) such that (x, y)  ϑ. (a)

(b)

Prove that if M is the projective limit of a diagram of finite discrete monoids, then M embeds as a closed submonoid of a product of finite monoids, and is therefore in particular compact and the congruences in ClpCon(M) separate points. Now, assume M is compact and that the clopen congruences separate points. Show that the sets of the form Kϑ,x0 := {x ∈ M | xϑx 0 },

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

308

(c)

8 Automata Theory

where ϑ ∈ ClpCon(M) and x 0 ∈ M, form a base of clopen sets for the topology on M. Let us write Q for the diagram whose shape is the poset ClpCon(M) with the inclusion order, and which sends ϑ to M/ϑ and an inclusion ϑ ⊆ ϑ  to the unique factorization of M  M/ϑ  through M  M/ϑ. Using the previous item, show that if M is compact and the congruences in ClpCon(M) separate points, then the collection of maps (M  M/ϑ)ϑ ∈ClpCon(M ) is a limiting cone over the diagram Q.

Exercise 8.2.9 Let A be a set and suppose that η : A → X and η  : A → X  are both free profinite monoids over A. This exercise shows that there is a unique iso-homeomorphism ϕ : X → X  which satisfies ϕ ◦ η = η . We use the results proved in Exercise 8.2.8. (a)

(b) (c) (d)

Let ϑ ∈ ClpCon(X  ) and write F := X /ϑ. Consider the function f : A → F defined by sending a ∈ A to [η  (a)]ϑ . By the universal property of X, extend f uniquely to a continuous homomorphism f: X → F. Prove that if ϕ : X → X  is a continuous homomorphism such that ϕ ◦ η = η , then we must have [ϕ(x)]ϑ = f(x) for all x ∈ X. Hint. Use the uniqueness of the extension f.  Show that there exists a continuous homomorphism ϕ : X → ϑ ∈ClpCon(X  ) X /ϑ. Show that ϕ is injective. Show that the image of ϕ is equal to the image of the embedding  X  → ϑ ∈ClpCon(X  ) X /ϑ.

Exercise 8.2.10 Let M be a profinite monoid and x ∈ M. (a) (b)

Show that the closure of {x n | n ∈ N ≥1 } is a commutative subsemigroup of M. Analogously to Proposition 8.39, prove that there exists a unique element y in the closure of {x n | n ∈ N ≥1 } such that yx = xy = x ω .

Exercise 8.2.11 Let M be a monoid. As in the case of A∗ , we say that a subset L of M is recognizable if there exists a homomorphism h : M → N with N a finite monoid and a subset P ⊆ N such that L = h−1 (P). We denote the set of recognizable subsets of M by Rec M . Show that the results for A∗ go through at this level of generality and thus, in particular, that: (a) (b)

Rec M is a Boolean residuation ideal in P (M). The dual space of the Boolean residuation algebra (Rec M , \, /) is the profinite completion of M. That is, the dual space admits a continuous

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

309

 is the projective limit monoid multiplication, and the resulting monoid M of the diagram of finite monoid quotients of M. The profinite completion of M may also be identified as the image of M under the left adjoint from monoids to topological monoids based on a Boolean space, that is, there is a  such that, for any topological monoid N monoid homomorphism η : M → M on a Boolean space and any homomorphism h : M → N, there exists a unique  → N such that  continuous homomorphism  h: M h ◦ η = h. Compare this with Remark 5.36 and Exercise 5.3.14; further see Gehrke (2016, Section 4). Note that Rec M need not be closed under concatenation product in P (M), as Exercise 8.2.12 shows. Exercise 8.2.12 This exercise uses the definition of a recognizable set in an arbitrary monoid as given in Exercise 8.2.11; it is based on Eilenberg (1974, Exercise III.12.9), also see Sakarovitch (2009, Exercise 2.6). (a) (b)

(c) (d)

Show that {0} is not recognizable in the monoid (Z, +). Let M be the structure (Z ∪ {e, x}, ·), where · is defined by k = k +  for k, l ∈ Z, xk = k x = k for all k ∈ Z, x 2 = 0, and em = me = m for all m ∈ M. Show that M is a monoid. Consider the equivalence relation on M defined by m ≡ n if, and only if m = n or m and n are both in Z. Show that ≡ is a congruence. Prove that L := {x} is recognizable in M but LL is not.

Exercise 8.2.13 Let h : L × L → L be a binary operation on a distributive lattice that is antitone in the first coordinate and preserves finite meets in the second coordinate. Prove that if F1, F2 ⊆ L are filters, then the set F := {b ∈ L | there exists a ∈ F1 such that h(a, b) ∈ F2 } is a filter in L. Use this to conclude that the set F defined at the end of the proof of Theorem 8.45 is a filter. 8.3 Equations, Subalgebras, and Profinite Monoids As follows from what we have seen in Sections 4.2 and 8.2, given a finite alphabet A, the set of all Boolean quotient spaces of the Boolean space underlying the free profinite monoid over A is in one-to-one correspondence with the set of Boolean subalgebras of Reg A. What is more is that these are all given by sets of what we called equations, which are really just pairs of elements of the dual space, see Definition 4.30 and Corollary 4.32. The interest of these equations is that, in order to describe a quotient space-subalgebra pair, one can use sets of equations that do

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

310

8 Automata Theory

not satisfy any special properties – in contrast with compatible equivalence relations (Definition 4.22 and Corollary 4.32), which are more difficult to understand. We recall in particular Example 4.35, which showed that in the presence of additional structure, by the use of equations, we may be able to obtain much smaller and simpler sets that characterize the subalgebras than the equivalence relations obtained from the full compatible preorders. This methodology has been greatly exploited in automata and formal language theory, and is in fact behind a great number of decidability results in the area. In this section, we want to give you a glimpse of the use of equations and inequations in language theory from a duality-theoretic perspective.

Profinite Monoid Inequations and Equations We first recall the definitions connecting sublattices and inequations as given in Section 4.2 (Definition 4.25), specialized to our use case here. For any element A∗ by L ∈ Reg A, we define a binary relation  L on  x  L y ⇐⇒ y ∈  L implies x ∈  L. A∗ . When L ⊆ Reg A, we also This is a preorder and  L is clopen as a subset of  define L .  L := L∈L

Remark 8.46 In the literature on automata and finite monoids, a syntactic preorder is commonly associated with a language L ⊆ A∗ . The classical definition (Schützenberger, 1955, Section 2), still in use today (Pin, 2021a, Section 1.4.4), is the following. For any L ⊆ A∗ and x, y ∈ A∗ , define the syntactic preorder  Ls of L by: x  Ls y if, and only if, for every u, v ∈ A∗ , if x ∈ u−1 Lv −1 , then y ∈ u−1 Lv −1 . This means that x  Ls y if, and only if, y  Q(L) x according to our definitions, that is,  Ls =( Q(L) ) op where we recall that Q(L) denotes the closure under quotienting of {L}. The original language L is then upward closed in the preorder  Ls , while it is downward closed in the preorder  L that we obtain from duality theory. As we have seen throughout this book (see Remarks 3.22 and 7.3), when applying duality theory to a different field, some conventions are bound to clash with the existing ones in the field, and one needs to choose whether to adapt or coexist. We here chose to keep the same definition of preorder associated to a sublattice as we did earlier in the book, which is thus the opposite of the definition of syntactic preorder. As it happens, this choice fits well with some of the existing literature on recognition by ordered monoids, in which the definition of syntactic preorder was also reversed, specifically some of the papers that use ordered monoids to analyze

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

311

the class of piecewise testable languages (Straubing and ThÃľrien, 1988; Pin, 1995; Henckell and Pin, 2000). Since piecewise testable languages will also be our focus later in this section, this gives another reason to adhere to the convention that we choose here. For any x, y ∈  A∗ , we define x  y := {L ∈ Reg A | if y ∈  L, then x ∈  L}, and note that this is a sublattice of Reg A. As in Definition 4.25, we also use the notation L |= x  y to mean that L ∈ x  y. If E is a set of pairs of  A∗ , we also define E := x  y. (x,y) ∈E

The content of Proposition 4.26, in this setting, is that we have an adjunction A∗ ×  A∗ ) op : − − : P (Reg A )  P ( 

(8.4)

whose fixed points on the left are sublattices of Reg A, and whose fixed points on the right are compatible preorders on  A∗ ; see Definition 4.22. Recall that, for a subset S of Reg A, Q(S) denotes the closure under quotienting of S. When E is a relation on  A∗ , let us denote by m(E) the biaction invariant closure of E, that is, the relation defined by m(E) := {(uxv, uyv) | (x, y) ∈ E, u, v ∈  A∗ }. We now prove that the operations of closure under quotienting and closure under the biaction are dual to each other, in the following sense. Proposition 8.47 For any set of pairs E ⊆  A∗ ×  A∗ and any subset S ⊆ Reg A, we have S ⊆ m(E) ⇐⇒ Q(S) ⊆ E. In particular, if E is closed under quotienting, then E = m(E). Proof We first show that, for any L ∈ Reg A and x, y ∈  A∗ , if Q(L) ⊆ x  y, ∗  then L |= uxv  uyv for every u, v ∈ A . Indeed, if Q(L) ⊆ x  y, then at least for every u, v ∈ A∗ , we have u−1 Lv −1 |= x  y, which means that L |= uxv  A∗ ×  A∗ → uyv. To show that this extends to all of  A∗ , consider the function f :   A∗ defined by f (u, v) := (uxv, uyv). This function is continuous because A∗ ×  the multiplication on  A∗ is continuous. Denote by C the inverse image of the set A∗ ) ∪ (  A∗ ×  L); this set is closed, and the assumption that Q(L) ⊆ x  y ( Lc ×  ∗ A∗ ×  A∗ , we get gives that A × A∗ is contained in C. Thus, since A∗ × A∗ is dense in 

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

312

8 Automata Theory

that C =  A∗ ×  A∗ , meaning that for any (u, v) ∈  A∗ ×  A∗ , we have L |= uxv  uyv, that is, L ∈ m({x  y}). This establishes the proposition in case S and E are singletons. For the general case, note that S ⊆ m(E) if, and only if, L ∈ m({x  y}) for every L ∈ S and every (x, y) ∈ E and Q(S) ⊆ E if, and only if, Q(L) ⊆ (x, y) for every L ∈ S and every (x, y) ∈ E. For the “in particular” statement, if E is closed under quotienting, then taking S := E in the equivalence gives E ⊆ m(E), and the other inclusion is clear because E ⊆ m(E) and the map − is antitone.  ∗

For x, y ∈  A∗ , let us write L |= x  y if, and only if, L ∈ m({x  y}), that is, L |= uxv  uyv for all u, v ∈  A∗ . We say in this case that L satisfies the profinite ∗

monoid inequation x  y, and we write ∗

∗

x  y := {L ∈ Reg A | L |= x  y}. ∗

Similarly, we write L |= x = y if, and only if, L |= uxv  uyv and L |= uyv  uxv for all u, v ∈  A∗ . In this case, we say that L satisfies the profinite monoid equation ∗ x = y. It follows in particular from Proposition 8.47 that, in the adjunction (8.4), the sublattices on the left that are closed under quotienting correspond exactly to the compatible preorders on the right that are monoid compatible, in the sense of the following definition. Note that this notion generalizes the definition of congruence, which is just a monoid-compatible equivalence relation. Definition 8.48 A binary relation R on a monoid M is monoid compatible if, for every u, v, x ∈ M, if uRv, then ux Rv x and xuRxv. We now show that profinite monoid equations for a regular language may be “tested” on the syntactic monoid. A similar result holds for profinite monoid inequations and ordered monoids (see Remark 8.46), but we do not fully develop that theory here. Recall that, when h : A∗ → M is a monoid homomorphism, with M a finite monoid, we denote by h¯ :  A∗ → M its unique continuous extension, which is also a homomorphism. Theorem 8.49 Let L ∈ Reg A and (x, y) ∈  A∗ ×  A∗ . The following are equivalent: (i) (ii) (iii) (iv)

∗

the language L satisfies the profinite monoid equation x = y; the (finite) Boolean residuation ideal B(L) is contained in x ≈ y; for the syntactic homomorphism h L : A∗  ML of L, we have h L (x) = h L (y); and there exists a homomorphism h : A∗ → M, with M a finite monoid, which ¯ ¯ recognizes L and such that h(x) = h(y).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

313

Proof By definition, (i) means that L ∈ m({x  y, y  x}). By Proposition 8.47, this is equivalent to Q({L}) ⊆ x ≈ y. Since x ≈ y is a Boolean subalgebra of Reg A and B(L) is the Boolean subalgebra of P ( A∗ ) generated by Q({L}) (Proposition 8.16), this gives that (i) is equivalent to (ii). For (ii) =⇒ (iii), recall that the languages recognized by ML are precisely those in B(L) (Corollary 8.20). Thus, in particular the language K := h−1 L ({h L (x)}) is  so (ii) gives that y ∈ K.  Using Exercise 8.2.7, we have that in B(L) and x ∈ K,  = h L −1 ({h L (x)}), so in particular h L (x) = h L (y), as required. K (iii) =⇒ (iv) is trivial, since h L recognizes L. For (iv) =⇒ (i), note that we have, for any u, v ∈  A∗ , ¯ ¯ h(x) ¯ h(v) ¯ ¯ h(y) ¯ h(v) ¯ ¯ h(uxv) = h(u) = h(u) = h(uyv), using the fact that h¯ is a homomorphism (Corollary 8.37). Using Exercise 8.2.7 again,  L = h¯ −1 (P), where P is a subset of M such that L = h−1 (P). Hence, for any u, v ∈  A∗ , we have ¯ ¯ xuv ∈  L ⇐⇒ h(uxv) ∈ P ⇐⇒ h(uyv) ∈ P ⇐⇒ x yv ∈  L, as required.



An important type of question in automata theory is, given a subclass C of the class of regular languages, whether the membership problem for C is decidable within the class of regular languages. More precisely, the C-membership problem asks for an effective procedure that allows one to decide, given a finite automaton, whether or not the language recognized by the automaton is in the class C. A powerful method, coming from the algebraic approach to language theory, is to give an equational criterion on the syntactic monoid of a regular language L for belonging to the given class C. Indeed, Proposition 8.47 shows that profinite monoid quotients of  A∗ correspond dually to the Boolean subalgebras of Reg A that are closed under quotienting. This is the kernel of Eilenberg’s theorem (Eilenberg, 1976) seen from a duality-theoretic point of view. Eilenberg considered pseudovarieties of regular languages, which are families V = (VA ) A∈Set f of Boolean subalgebras of Reg A closed under quotienting, satisfying the further property that the Boolean algebras VA are stable under inverse image by homomorphisms between free monoids. An extension of Proposition 8.47 then shows that the class of syntactic monoids of languages in V is definable by profinite equations that are moreover invariant under substitutions. Such classes of finite monoids are called pseudovarieties of finite monoids, and, by a finite version of Birkhoff’s theorem, are exactly those classes of finite monoids closed under homomorphic images, submonoids, and finite products. Eilenberg’s theorem

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

314

8 Automata Theory

can thus be viewed as a duality between pseudovarieties of regular languages and pseudovarieties of finite monoids; for further details, see Gehrke (2016). Rather than exposing this general theory any further here, we illustrate the method by working out a particular case, namely that of piecewise testable languages. In this case, a decidable criterion for membership was first obtained in Simon (1975), who showed that the piecewise testable languages are exactly the languages recognized by finite J -trivial monoids. Our aim in the rest of this section is to obtain a profinite equational characterization of piecewise testable languages by dualitytheoretic means and we will explain how this shows that the class has a decidable membership problem (Corollary 8.68). Positively Piecewise Testable Languages In order to define the positively piecewise testable and piecewise testable languages, we need the subword ordering. To this end, it is convenient to think of a finite word w ∈ A∗ as a function |w| → A, where |w| is identified with the totally ordered set 1 < 2 < · · · < |w| and the function sends i to the i th letter of w. In this context, we often also write w(i) for the i th letter of w. We say that u ∈ A∗ is a (scattered) subword of w if there exists an order embedding ϕ from |u| to |w| such that w(ϕ(i)) = u(i) for every i ∈ |u|, that is, if u is a subsequence of w. For example, the word abbc is a subword of acabbac, but not of acaba, and not of abc either. We will write u + w if u is a subword of w, and we note that + defines a partial order on A∗ . Definition 8.50 Let A be a finite alphabet and L ⊆ A∗ a language. A language is called positively piecewise testable 5 if it is a finitely generated up-set in the partial order ( A∗, +) and piecewise testable if it is a Boolean combination of positively piecewise testable languages. We denote by PT+A the lattice of positively piecewise testable languages and by PT A the Boolean algebra of piecewise testable languages. For the rest of this section, we fix a finite alphabet A and write PT+ and PT for PT+A and PT A, respectively. We also write Reg for Reg A. In formal language theory, a language class is often seen as a fibered collection of lattices, or Boolean algebras, where the alphabet may vary and morphisms between alphabets induce morphisms between the lattices. We do not need this structure in this section, and we only remark here that it is related to the fibrational approach via hyperdoctrines that we will point to at the end of Section 8.4. Any positively piecewise testable language is regular, since it is easy to check with a finite automaton whether any of a finite number of subwords appear in a word (see Exercise 8.3.1). It follows that any piecewise testable language is regular, since Reg is a Boolean algebra (Proposition 8.24). 5

What we call a positively piecewise testable language is also known as a shuffle ideal (Pin, 1995) or as a language in the (marked) polynomial closure of {∅, A∗ } (Pin, 2011).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

315

Toward the equational characterization of piecewise testable languages, we first show how the collection PT+ may be identified with an inequational property on the free profinite monoid. To this end, first note that the collection PT+ of finitely generated up-sets of ( A∗, +) is a sublattice of P ( A∗ ). By Exercise 7.2.3 this statement is equivalent to the statement that ( A∗, +) is finitely MUB-complete in the sense of Definition 7.50, that is, for any finite set F ⊆ A∗ , there is a finite  set G ⊆ A∗ such that ↑G = w ∈F ↑w. The latter is indeed the case, but in fact something much stronger is true. Even though ( A∗, +) is not bifinite, it is what is known as a well-quasi-order, that is, every up-set of ( A∗, +) is finitely generated. Indeed, this was shown early on by Higman (1952). For completeness, we include a proof. Note that while the following proof uses the axiom of choice, one can also give a constructive proof, for example, by induction on the size of the alphabet, see Murthy and Russell (1990). Theorem 8.51 (Higman, 1952, Theorem 4.4) For any finite alphabet A, the poset ( A∗, +) is a well-quasi-order. Proof Let S = {U ∈ U ( A∗, +) | min(U) is infinite}. If S is empty, then we are

done. Suppose S  ∅. Let C be a chain in S. We will show that C is in S. Clearly,

C is an up-set. Notice that every up-set is the up-set of its minimal elements

since the down-set of any word is finite. Suppose that the set F := min( C) is finite. Then, since C is a chain, there is a single V ∈ C with F ⊆ V . But then

V ⊆ C = ↑F ⊆ V , which is a contradiction since V ∈ S. So, C ∈ S and by Zorn’s lemma (Lemma 2.7), there is a maximal element U ∈ S. We now have  min(U) = {min(U) ∩ a A∗ | a ∈ A} and thus Ua = min(U)∩a A∗ = aVa is infinite for some a ∈ A. Now, let W = U∪↑Va . Then, Va is infinite and each element in Va is minimal in W . To see this, notice that if w ∈ Va , then aw ∈ min(U) and thus w  U. Further, if v ∈ Va and v + w, then av + aw and av, aw ∈ min(U), so v = w. Thus, W ∈ S and Va  U so that U  W , which contradicts the maximality of U.  Since PT+ is the lattice of up-sets of a well-quasi-order, we note that we are in a special case of the spectral domains considered in Section 7.2. Thus, by Corollary 7.48 and Exercise 7.2.4, the spectral space dual to PT+ is isomorphic to the poset Idl( A∗ ) in the Scott topology. Since PT+ is a sublattice of the Boolean algebra of regular languages, its dual space Idl( A∗, +) can also be characterized by inequations on the dual space  A∗ of Reg. Note also that PT+ is closed under quotienting. Indeed, for any words u, v ∈ A∗ and L ∈ PT+ , we have u−1 Lv −1 ∈ PT+ : if uwv ∈ L and w + w , then uwv + uw  v, so uw  v ∈ L as well. Thus, in order to obtain a characterization of PT+ via profinite

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

316

8 Automata Theory

inequations, Proposition 8.47 implies that we may look for a characterization via profinite monoid inequations. The lattice PT+ contains exactly the regular languages that ∗ A∗ . satisfy the profinite monoid inequation x   for every x ∈ 

Proposition 8.52

Proof Let us first show that any language L in PT+ satisfies the profinite monoid equation. Since PT+ is closed under quotienting, it suffices to show that PT+ ⊆ x   , by Proposition 8.47. But if L is in PT+ , then  ∈ L clearly implies L = A∗ , L= A∗ , and thus x ∈  L for every x ∈  A∗ . since  + w for every w ∈ A∗ , so that  ∗ Conversely, suppose that L |= x   for every x ∈  A∗ . Suppose that w + w  and w ∈ L. By induction on the difference in length between w  and w, it suffices to treat the case where w = uv and w  = uav for some u, v ∈ A∗ and a ∈ A. Now, ∗ since L |= a  , we also get L |= uav  uv. Therefore, since uv ∈ L, we have uav ∈ L.  Propositions 8.47 and 8.52 now yield the following. Theorem 8.53 The Priestley dual space of PT+ is homeomorphic to the quotient space  A∗ /, where  is the smallest compatible preorder containing the profinite ∗ A∗ . monoid inequations x   for every x ∈  Note that this result in particular gives an equational way of looking at the spectral domain Idl( A∗ ) as follows. As noted above, the spectral space dual to PT+ can also be described as Idl( A∗ ) equipped with the Scott topology. Thus, using Theorem 6.4, we also get the following. Corollary 8.54 The Priestley dual space of PT+ is homeomorphic to (Idl( A∗ ), σ p, ⊇), where σ p := σ ∨ σ ∂ and σ is the Scott topology on (Idl( A∗ ), +). We further have a canonical map π :  A∗  Idl( A∗ ), dual to the inclusion + PT → Reg, given by continuously extending the inclusion A∗ → Idl( A∗ ) that A∗ to sends w ∈ A∗ to the principal ideal ↑w. Concretely, this map sends u ∈  ∗  (see Exercise 8.3.3). Theorem 8.53 implies that the ideal {w ∈ A | u ∈ ↑w}

π(u) ≤ π(v) if, and only if, u PT+ v. Defining an appropriate monoid multiplication on Idl( A∗ ), the map π becomes a continuous monoid homomorphism (again see Exercise 8.3.3 for more details).

Piecewise Testable Languages Now, moving towards a first characterization of the dual space of PT, recall from Proposition 3.34 and Corollary 3.35 that, since PT is the Boolean envelope of PT+ ,

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

317

the Stone dual space of PT is simply given by forgetting the order of the Priestley dual space of PT+ . We thus also get the following result. Corollary 8.55 The dual space of PT is homeomorphic to  A∗ /( ∩ !), where  is the preorder of Theorem 8.53. Remark 8.56 It was recently proved that the dual space of the Boolean algebra of piecewise testable languages over an alphabet A with | A| = n is homeomorphic to n−1 the ordinal ωω + 1 with the interval topology (Pouzet, 2023, Theorem 11), based on earlier, more general results by Bekkali et al. (2007). While Corolary 8.54 tells us something about what the dual space of PT+ looks like, Corollary 8.55 and Theorem 8.53 tell us something about the (in)equations satisfied by PT+ and PT, and the latter type of information is what can lead to a decidable characterization. In order to obtain such a decidable equational characterization for piecewise testable languages, we need a better description of the congruence  ∩ ! of Corollary 8.55. Let us introduce a notation for this congruence. A∗ Definition 8.57 For A a finite alphabet, we write ≈PT for the congruence on  dual to the subalgebra PT ⊆ Reg. We first identify two properties of this congruence that will be important for this description. These properties take the form of quasi-identities, which is the term commonly used in universal algebra for an expression of the form   if s1 = t 1 and · · · and s n = t n , then (s0 = t 0 ) , where the si and t i are terms. Lemma 8.58 Let  be a monoid-compatible preorder on a monoid M such that x  1 for every x ∈ M and denote its associated equivalence relation by ≈. Then, for any x, y, u ∈ M we have (a) (b)

if uxy ≈ u, then ux ≈ u, and if yxu ≈ u, then xu ≈ u.

Proof Let x, y, u ∈ M. Note that, since  is monoid-compatible and x  1 for every x ∈ M, we also get ux  u · 1 = u and uxy  ux · 1 = ux. Now, for the first item, suppose that uxy ≈ u. Then u ≈ uxy  ux  u, so ux ≈ u. The proof of the second item is symmetric.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press



318

8 Automata Theory

Using Theorem 8.53, the congruence ≈PT dual to PT satisfies both of the properties of Lemma 8.58. We now aim to show that these properties are equivalent to profinite monoid equations, and that the topological monoid congruence ≈PT is in fact generated by those equations. This will yield a proof of Simon’s theorem and, as we will see, the decidability of the membership problem for piecewise testable languages. Let us first show how the properties of Lemma 8.58 can be written as profinite monoid equations. Here, the operations ω and ω −1 that we introduced in Definition 8.40 and immediately after (see also Exercise 8.2.10) are important, as they allow us to construct profinite terms. Lemma 8.59 Let ≈ be a closed congruence on a profinite monoid M. (a)

The following are equivalent: (i) (ii)

(b)

For any u, x, y ∈ M, if uxy ≈ u, then ux ≈ u. For any x, y ∈ M, (x y) ω x ≈ (x y) ω .

The following are equivalent: (i) (ii)

For any u, x, y ∈ M, if yxu ≈ u, then xu ≈ u. For any x, y ∈ M, x(yx) ω ≈ (yx) ω .

Moreover, if any of these properties hold, then z ω+1 ≈ z ω for any z ∈ M. Proof We only prove that (i) and (ii) in (a) are equivalent, and imply that z ω+1 ≈ z ω , the proof for (b) is symmetric. Suppose (i) holds. We first show that z ω+1 ≈ z ω for any z ∈ X. Indeed, applying (i) with u = z ω , x = z, and y = z ω−1 , we have uxy = z ω z ω = u, so that z ω+1 ≈ z ω . In particular, (xy) ω xy ≈ (xy) ω , so that (i) now gives (x y) ω x ≈ (x y) ω . Conversely, suppose (ii) holds and suppose that uxy ≈ u. Then, for any n, we have u ≈ u(x y) n , so, since ≈ is closed, u ≈ u(xy) ω . Thus, ux ≈ u(x y) ω x ≈ u(x y) ω ≈ u.



Remark 8.60 The two profinite equations (a)(ii) and (b)(ii) are known in the literature as the profinite equations for R-trivial and L-trivial monoids, respectively, and it is known that together they characterize the class of J -trivial finite monoids. The equation z ω ≈ z ω+1 that appeared in Lemma 8.59 characterizes the class of aperiodic or H -trivial monoids. The letters L, R, J , and H refer to Green’s equivalence relations on monoids, and the adjective “trivial” asserts that these relations are equal to the diagonal in the profinite monoid under consideration. We will not need to consider these relations in any more detail here, and our development is independent of any of the facts mentioned in this remark. We refer the reader who wants to know more about Green’s relations to Exercise 8.3.2, and standard texts

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

319

on monoid theory, such as for example Eilenberg (1974, 1976), Almeida (1995), Rhodes and Steinberg (2008), and Pin (2022). We here make a slightly non-standard definition of J -trivial profinite monoid which however coincides with the usual definition, see Exercise 8.3.2. A∗ Definition 8.61 Denote by ≈ J the smallest profinite monoid congruence on  ∗ that contains, for every x, y ∈  A , the pairs ((x y) ω x, (x y) ω ) and (x(yx) ω, (yx) ω ) . Denote the residuation ideal of Reg A dual to ≈ J by Jtriv . We call a profinite monoid M J -trivial provided for all x, y ∈ M, we have (xy) ω x = (xy) ω and x(yx) ω = (yx) ω , and we call the elements of Jtriv the J -trivial languages of Reg. By Theorem 8.49, a regular language L is J -trivial if, and only if, its syntactic monoid is J -trivial if, and only if, it is recognized by some J -trivial monoid. Our final aim in this section is to prove that PT = Jtriv or equivalently that ≈ J = ≈PT . The inclusion PT ⊆ Jtriv actually follows immediately from Lemmas 8.58 and 8.59: Corollary 8.62 For any x, y ∈  A∗ , we have (x y) ω x ≈PT (x y) ω and x(yx) ω ≈PT (yx) ω . That is, ≈ J ⊆ ≈PT and dually PT ⊆ Jtriv . It now remains to show that J triv ⊆ Jtriv . Since PT is defined as the Boolean algebra generated by the lattice PT+ = U ( A∗, +), we are in the situation that we discussed in Section 4.7. Theorem 4.79 therefore gives that a language L is in PT if, and only if, it has bounded alternation height in the subword order +. We will prove that any language L that satisfies the profinite monoid equations defining ≈ J (and thus the quasi-identities of Lemma 8.58) has bounded alternation height. For this final step, we need to develop a small amount of finite monoid theory, which gives just a flavor of this rich field. Before giving the formal details of the proof, we first give an intuition. Let w = a1 · · · ak be a finite word and let (−) : A∗ → M be a homomorphism to a finite monoid. We will analyze how w, i.e., the value of w under the homomorphism M can be computed “from left to right;” a similar analysis applies by computing the value “from right to left.” Reading w from left to right, the value w can be computed, starting from 1 M , by first computing a1 , then a1 a2 , then a1 a2 a3 , and so on. In this computation, which takes k steps, only certain steps will change the value; positions where this happens will be called “right unstable” below. By definition, only the unstable positions of a word contribute to its value, and we may thus reduce a word w to obtain a subword

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

320

8 Automata Theory

r (w) that only contains its right unstable positions. The idea of the rest of the proof is that, if M is a finite monoid that satisfies the quasi-identities, then the number of reduced words is bounded. This will give a bound on the alternation height as a subset of the poset ( A∗, +) of any regular language L that is recognized by M. Now, more formally, we make the following definitions. A reader familiar with finite monoid theory may recognize the similarity of the reduction notion that we introduce here with the Karnofsky–Rhodes expansion, see, for example, Rhodes et al. (2022, Section 2). Definition 8.63 Let M be a finite monoid and (−) : A∗ → M a homomorphism; we write w for the image of w ∈ A∗ under the homomorphism. For any finite word w = a1 · · · ak ∈ A∗ with each ai ∈ A and for each 0 ≤ i ≤ k, we write wi for the length i prefix of w, that is, w0 :=  and wi := wi−1 ai for all 1 ≤ i ≤ k. We will say a position i ∈ {1, . . . , k} is (right) stable if wi = wi−1 , and (right) unstable otherwise. Let t 1 < · · · < t  be an enumeration of the unstable positions of w. We define r (w) := at1 · · · at , the subword of w on the unstable positions. A simple induction on the number of unstable positions in w shows that the words w and r (w) have the same value, that is, w = r (w). We call r (w) the right reduction of w and we call a word in A∗ right reduced if all of its positions are unstable, that is, if w = r (w). The notions of left (un)stable, left reduction, and left reduced are defined analogously, using the sequence of suffixes of w. We write (w) for the left reduction of w. We begin by showing that the quasi-identity identified in Lemma 8.58(a) implies that, if M is finite, then there are finitely many right-reduced words for any homomorphism from A∗ to M. We only state and prove the direction that we need in the proof, however, the converse is also true (see Exercise 8.3.4). Lemma 8.64 Let (−) : A∗ → M be a homomorphism to a finite J -trivial monoid M. Then, there are finitely many right-reduced words for the homomorphism (−). Proof We show that, for any right-reduced word w = a1 · · · ak in A∗ of length k, and for any 1 ≤ i < j ≤ k, the prefixes wi and w j have distinct values. This will then imply in particular that a right-reduced word can only have length at most |M |, and thus there are finitely many right-reduced words. Let w = a1 · · · ak ∈ A∗ be a right-reduced word of length k, and suppose, towards a contradiction, that there exist 1 ≤ i < j ≤ k such that wi = w j . This means that a1 · · · ai ai+1 · · · a j = a1 · · · ai . Thus, applying the quasi-identity of Lemma 8.59(a)(i) with u := a1 · · · ai , x := ai+1 , y := ai+2 · · · a j , we see that uxy = u, so ux = u. But this means that wi = wi+1 , so that i + 1 is a right-stable point, contradicting the assumption that w is right reduced. 

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

321

Lemma 8.65 Let (−) : A∗ → M be a homomorphism to a finite monoid M, w ∈ A∗ , a ∈ A, and suppose that the word r (w)a is a subword of w. Then, wa = w. Proof Let w ∈ A∗ , a ∈ A and suppose that the word r (w)a is a subword of w. If r (w) = ε, then 1 M = b = w for all b ∈ A occurring in w and thus, in particular, wa = w · a = 1 M = w. Now, suppose w = a1 · · · ak with each ai ∈ A and let t be an unstable position in w. Further, let t  be the last unstable position preceding t. If t is the first unstable position in w, we take t  = 0. We claim that the letter at cannot appear anywhere strictly between positions t  and t. To see this, notice that, since for every t  < j < t, j is stable, we have wt  = w j = w j−1 a j = wt  a j , while wt = wt−1 at = wt  at  wt  . Thus, we must have a j  at , as claimed. These remarks imply that, if ϕ is an embedding of r (w) = at1 · · · at as a subword of w, then for every 1 ≤ i ≤ , ϕ(i) ≥ t i . Finally, let ψ be an embedding of r (w)a as a subword of w. Then ψ must send the last position to some position j with j > t  . Then, since all positions after t  are stable, we get in particular that w j−1 = w j = w, so wa = w j−1 a = w j = w.



Applying the exact same reasoning when starting to read w from right to left, for the left reduction (w) of w, we have, analogously to Lemma 8.65, that a(w) + w implies aw = w, and, analogously to Lemma 8.64, that the quasi-identity yxu = u ⇒ xu = u implies that there are finitely many left-reduced words. Proposition 8.66 Let M be a finite J -trivial monoid, (−) : A∗ → M a homomorphism, and L a regular language recognized by (−). Then, L is piecewise testable. Proof As explained above, by Theorem 4.79, it suffices to prove that L has bounded alternation height in the poset ( A∗, +). The images of the reduction functions r : A∗ → A∗ and  : A∗ → A∗ obtained from the homomorphism (−) are finite by Lemma 8.64 and its symmetric “left” version. Therefore, the set F := {uav | u ∈ im(r), a ∈ A, v ∈ im()} is also finite. For any word w ∈ A∗ , denote by f (w) the set of subwords of w that are in F. Clearly, if w + w , then f (w) ⊆ f (w  ). Claim Suppose w + w . If w  w , then f (w)  f (w  ). Proof of Claim. Let w + w . Then, f (w) ⊆ f (w  ). We prove the claim by contraposition. To this end, suppose that f (w) = f (w  ). Since w  can be obtained from w by inserting a finite number of letters, we can assume w = uv and w  = uav for some u, v ∈ A∗ and a ∈ A; if we prove for this special case that w = w , then the

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

322

8 Automata Theory

contrapositive of the claim follows by an easy induction on the length difference between w  and w. Now, since uav = w , we have r (u)a(v) + w , so r (u)a(v) ∈ f (w  ). Since f (w) = f (w  ), we also have that r (u)a(v) ∈ f (w), so r (u)a(v) is a subword of w = uv. This implies that either r (u)a + u or a(v) + v, according to where the embedding sends the “middle” position a. Let us assume r (u)a + u; the other case is symmetric. Lemma 8.65 gives that ua = u, and thus w = uv = uav = w , and the claim is proved. Finally, let x 1 + y1 · · · + x n + yn be an alternating chain for L, and let P be the image of L by (−). Since L is recognized by (−), we have w ∈ L if, and only if, w ∈ P. Since x 1 + y1 · · · + x n + yn alternates in and out of L, no two consecutive elements of the chain can have the same image by (−). Therefore, by the claim, we have f (x 1 )  f (y1 ) · · ·  f (x n )  f (yn ). Since each inclusion in this chain is strict, the cardinality of the set increases by at least 1, so it follows that 2n ≤ |F | + 1, giving a bound on the length n of the alternating chain.  Putting the above results together, we get that the profinite equations defining ≈ J characterize the congruence ≈PT dual to PT. Theorem 8.67 The Boolean residuation algebra PT of piecewise testable languages is dual to the smallest profinite monoid congruence ≈ on  A∗ that contains ω ω ω ω ∗  ((x y) x, (xy) ) and (x(yx) , (yx) ) for every x, y ∈ A , that is, ∗

∗

PT = (x y) ω x = (xy) ω, x(yx) ω = (yx) ω . Proof Recall that ≈ J is by definition the smallest profinite monoid congruence on  A∗ that contains the equations, and ≈PT is the congruence dual to PT; we show they are equal. By Corollary 8.62, ≈ J is contained in ≈PT . On the other hand, if L is a regular language satisfying the equations defining ≈ J , then Proposition 8.66 together with Lemma 8.59 imply that L is piecewise testable. Thus, the Boolean algebra associated with ≈ J is contained in PT, so that ≈PT is contained in ≈ J .  It follows in particular that the membership problem for piecewise testable languages is decidable. Indeed, by Theorem 8.22, the syntactic monoid of a regular language L is computable from any automaton recognizing it, and Theorem 8.49, applied to the profinite monoid equations defining ≈ J , shows that these equations may be tested on the syntactic monoid. We conclude the following. Corollary 8.68 A regular language L is piecewise testable if, and only if, its syntactic monoid ML satisfies the equations (x y) ω x = (x y) ω and x(yx) ω = (yx) ω

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.3 Equations, Subalgebras, and Profinite Monoids

323

for every x, y ∈ ML . In particular, it is decidable for a given regular language whether it is piecewise testable. Exercises for Section 8.3 Exercise 8.3.1 Let A be a finite alphabet and w ∈ A∗ . Construct a finite automaton that recognizes the language ↑w := {v ∈ A∗ | w + v}. Deduce that any finitely generated up-set (and thus, by Theorem 8.51, any up-set) of ( A∗, +) is regular. Exercise 8.3.2 Let M be a monoid. For any m, n ∈ M, we define def

m  L n ⇐⇒ there exists x ∈ M such that xn = m, def

m  R n ⇐⇒ there exists y ∈ M such that ny = m, and def

m  J n ⇐⇒ there exist x, y ∈ M such that xny = m. (a)

Prove that  L,  R ,  J are preorders on M and that  L ∪  R ⊆  J .

The equivalence relations associated to these preorders are denoted L, R, J , respectively. In general, a monoid is called L-trivial, R-trivial, or J -trivial if the corresponding equivalence relation is equal to the diagonal (also known as “trivial”) relation. We show in this exercise that this definition of J -trivial coincides with the one given in the main text for profinite monoids, and also that J -trivial is equivalent to the conjunction of L-trivial and R-trivial for profinite monoids. (b)

Prove that if J is the trivial relation, then L and R are also both the trivial relation.

Now, suppose that M is a profinite monoid. (c) (d)

Prove that, for any x, y ∈ M, (x y) ω = x(yx) ω−1 y. Prove that, for any x, y ∈ M, (x y) ω xR (x y) ω and x(yx) ω L(yx) ω .

(e)

(f)

Conclude that if R is the trivial relation then the equation (xy) ω x = (xy) ω holds in M, and if L is the trivial relation, then the equation x(yx) ω = (yx) ω holds in M. Conversely, prove that if both (x y) ω x = (xy) ω and x(yx) ω = (yx) ω for all x, y ∈ M, then J is the trivial relation. Hint. Show first that, if mxy = m in M, then m(x y) ω = m. Then, generalize this idea and use the given equations. Conclude that, in any profinite monoid M, the following are equivalent: (i) the equivalence relation J is trivial;

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

324

8 Automata Theory

(ii) (iii)

the equivalence relations L and R are both trivial; and the equations (x y) ω x = (x y) ω and x(yx) ω = (yx) ω hold for all x, y ∈ M.

Exercise 8.3.3 Recall that Idl( A∗, +) is a spectral domain in the Scott topology, whose compact-opens are the sets of the form K F := {I ∈ Idl( A∗ ) | I ∩ F  ∅}, A∗ → Idl( A∗ ) was with F a finite subset of A∗ . Also, recall that a function π :   defined by π(u) := {w ∈ A∗ | u ∈ ↑w}. (a) (b)

(c)

Prove that π is well defined, and that in particular for any u ∈ A∗ , π(u) = ↓u, the principal down-set of u. Prove that π is continuous with respect to the patch topology on Idl( A∗ ). Hint. A∗ for every finite subset F of It suffices to prove that π −1 (K F ) is clopen in  ∗ A. Prove that there is a well-defined continuous monoid multiplication on Idl( A∗ ) given, for I, J ∈ Idl( A∗ ), by I · J := {ww  | w ∈ I, w  ∈ J}.

(d) (e)

Prove that π(uv) = π(u) · π(v) for any u, v ∈ A∗ . Conclude, using Exercise 8.2.6, that π is a continuous monoid homomorphism  A∗ → Idl( A∗ ).

Exercise 8.3.4 Let (−) : A∗ → M be a monoid homomorphism and suppose that there exist u, x, y ∈ A∗ such that uxy = u but ux  u. Prove that there exist infinitely many right-reduced words for the homomorphism A∗ → M. Exercise 8.3.5 Prove that the complement of the language L of Exercise 8.1.2 is positively piecewise testable. Exercise 8.3.6 In this exercise we let A = {a, b} and consider the language L := {w ∈ A∗ | |w|a > |w|b }, where, for w ∈ A∗ and x ∈ A, we write |w|x for the number of occurrences of the letter x in the word w. The language L is often called “majority” in the literature since it contains the set of words in which the majority of letters are a. For any k ∈ Z, define the language L k := {w ∈ A∗ | |w|a − |w|b > k}. (a)

Show that the closure under quotienting Q(L) of L is equal to {L k | k ∈ Z}. Conclude in particular that L is not regular.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.4 Open Multiplication

(b)

(c)

(d)

(e)

(f)

325

Show that Z is, up to isomorphism, the discrete dual of the complete Boolean algebra generated by Q(L) and that the monoid morphism {a, b}∗ → Z, w → |w|a − |w|b is the discrete dual of the inclusion of this algebra in the powerset of {a, b}∗ . Verify that the kernel of this monoid morphism is the syntactic congruence ≡ L and thus that the morphism is, up to isomorphism, the syntactic morphism of L. Deduce that the quotienting Boolean algebra B(L) generated by L is equal to the Boolean algebra generated by the languages L k . Show that it is not a complete Boolean algebra. Prove that B(L) is isomorphic to the Boolean subalgebra M of P (Z), considered in Example 4.35, consisting of the subsets S of Z such that both S ∩ Z+ is finite or co-finite and S ∩ Z− is finite or co-finite. Here, as in Example 4.35, Z+ denotes the set of positive integers and Z− denotes the set of negative integers. Write R+ for the graph of the addition operation on Z, that is, R+ = {(x, y, z) ∈ 3 Z3 | x + y = z}, and denote by R+ the closure of R+ in the space (Z+∞ −∞ ) . +∞ Show that, for any u ∈ Z−∞ , the elements (−∞, +∞, u) and (+∞, −∞, u) are in R+ ; that is, R+ contains the set {(−∞, +∞), (+∞, −∞)} × Z+∞ −∞ . Conclude that extends the addition that there is no continuous binary function on Z+∞ −∞ on Z. Using Example 4.35, show that B(L) is given by the following set of spatial equations on β A∗ : { μx ≈ μ, x μ ≈ μ | μ ∈ β A∗, x ∈ A}. Here, μx and x μ denote the left and right action of A∗ on β A∗ induced by the biaction dual to the quotienting biaction of A∗ on P ( A∗ ). 8.4 Open Multiplication

What better to end with than something open? This book is just a small sampler of the things you can do with duality – even within our two chosen applications of domain theory and automata theory, the theory goes much beyond what we have touched on here. In Sections 8.2 and 8.3, we saw that the multiplication of a profinite monoid can be seen as dual to residuation structure on the dual Boolean algebra, or as dual to a biaction structure by quotienting operations. In this final section, we consider a third possibility for seeing the multiplication of a profinite monoid as the dual of a binary operation on the dual Boolean algebra, which applies only in the special case where multiplication is an open map. Theorem 8.45 establishes a connection between the residual operations (\, /) on a Boolean algebra and continuous binary operations f on its dual Boolean space. One may wonder what it takes for the forward image map given by the binary

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

326

8 Automata Theory

operation to be the dual of an operation. Using Definition 4.38, we obtain the following requirements on f : (a) (b)

For all x ∈ X the inverse image f −1 (x) is closed. The forward image of a pair of clopens is clopen.

Without assuming that f is also continuous, these are not very natural conditions on a map between topological spaces. However, we do obtain the following useful corollary. Corollary 8.69 Let (X, f ) be a binary topological algebra based on a Boolean space and let B be the dual Boolean algebra of clopen sets. The operation f is an open mapping if, and only if, B is closed under the forward operation (U, V ) → U • V := { f (u, v) | u ∈ U, v ∈ V }, that is, if, and only if, the complex multiplication of two clopen sets of X is again clopen. In this case, the graph of f is the relational dual to the operation • on B. Proof Since all continuous maps from compact spaces to Hausdorff spaces are closed mappings, it follows that B is closed under the complex operation • if, and only if, f is an open map. The conditions required for the graph R of f to be the dual of this operation • are that f −1 (x) = R[_, _, x] is closed for each x ∈ X and that U • V = R[U, V, _] is clopen whenever both U and V are clopen. For f continuous and X Hausdorff, the first condition always holds. The second condition holds if f is an open mapping.  Note that, in the special case of languages and the free profinite monoid, a subtlety arises from the mixing of discrete and topological duality that happens there. More precisely, if B is a Boolean algebra of regular languages, then it is possible that B fails to be closed under the multiplication in P ( A∗ ), even though its natural embedding in the powerset of the dual space X of B is a Boolean subalgebra closed under complex multiplication. That is, the Boolean subalgebra { L | L ∈ B} of P (X ) may be closed under complex multiplication, even if B itself is not, see Example 8.71. Example 8.70 One of the, if not the, most famous theorems about regular languages is Kleene’s theorem (Kleene, 1956) which asserts that Reg A is closed under concatenation product in P ( A∗ ). This fact also has a very natural duality proof in categorical logic (Marquès, 2021). Example 8.71 A free profinite group, or in fact any topological group, has open multiplication. Indeed, for any element g of a topological group G, the multiplication x → xg is a homeomorphism of G with inverse x → xg−1 , and it is thus in particular

an open map. Therefore, for any U, V ⊆ G with U open, U • V = g ∈V Ug is open.

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.4 Open Multiplication

327

However, the product of group languages is not necessarily a group language. Indeed, let L := (a2 ) ∗ a. Then L is recognized by the finite group Z2 , but it is not difficult to see that LL = (a2 ) + can not be recognized by a finite group (see Exercise 8.4.1). We refer the reader to Ribes and Zalesskii (2010) for an extensive survey of the theory of profinite groups. We do have the following sufficient condition for an algebra of regular languages to be closed under concatenation product, in the situation of Corollary 8.69. Proposition 8.72 Let f :  A∗  X be a profinite monoid quotient, B be the Boolean ( ) the Stone embedding of B in P (X ), and assume subalgebra of Reg A dual to X,  that the multiplication on X is an open mapping. Then, for L, K ∈ B, the element  is the least N ∈ B for which LK ⊆ N M ∈ B corresponding to the clopen  L•K ∗ holds in P ( A ). If in addition {u} is in B for every u ∈ A∗ , then B is closed under concatenation product in P ( A∗ ). Proof Since the multiplication is open in X, for every L, K ∈ B, there is an M ∈ B  = M.  We first show that, even if LK  B, this entails that M is the least with  L•K element of B above LK in P ( A∗ ). Let N ∈ B. Then, we have LK ⊆ N ⇐⇒ L ⊆ N/K  = N/  K  ⇐⇒  L ⊆ N/K = ⊆N  ⇐⇒ M L•K ⇐⇒ M ⊆ N . Finally, suppose that {u} ∈ B for every u ∈ A∗ , then also {u} c ∈ B, and thus u  LK implies LK ⊆ {u} c , which implies M ⊆ {u} c , and thus u  M. By contraposition, we have M ⊆ LK and thus LK = M.  Openness of the multiplication has been studied in language theory, but it also appears in categorical first-order logic. Classical research in categorical first-order logic is concerned with extending Stone duality to a setting where the Boolean algebras, such as Makkai (1987), are equipped with quantifiers, a topic that is also still under development today, see, for example, Lurie (2019) or van Gool and Marquès (2024). A fundamental structure in that field is that of a Boolean hyperdoctrine, which is a functor P valued in Boolean algebras which is required to satisfy two algebraic conditions on the morphisms, called the Beck–Chevalley and Frobenius conditions, both of which have a natural meaning through duality theory. In one recent result in this direction (Marquès, 2021), openness is combined with equidivisibility, a notion which has been considered in language theory (Almeida and Costa, 2009). A semigroup S is called equidivisible if, for all elements u1, u2, v1, v2 ∈ S such that u1 u2 = v1 v2 , there exists an element k ∈ S such

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

328

8 Automata Theory

that either u1 k = v1 and u2 = kv2 , or v1 k = u1 and v2 = ku2 . Relative to duality theory, this condition on the binary operation · to be equidivisible is precisely the “back” condition of bounded morphism (Definition 4.54). In Marquès (2021, Theorem 4.3), it is shown that a profinite semigroup S is dual to a first-order theory enriching the theory of bounded linear orders if, and only if, the multiplication on S is open and equidivisible. To obtain this result, it is shown that, under Stone duality, the openness of the multiplication is dual to the Beck–Chevalley condition, and the equidivisibility corresponds to the Frobenius condition. This theorem thus establishes yet another connection between language theory and topological algebra, mediated by extended Stone duality.

Exercises for Section 8.4 Exercise 8.4.1 Let A = {a} and L ⊆ A∗ the language of words of odd length. (a) (b)

(c)

Show that the syntactic monoid of L is the two-element group Z2 , so that L is a group language. Show that the syntactic monoid of L · L, the language of words of non-zero even length, is the three-element monoid (Z2 ) I , that is, the monoid obtained from Z2 by adjoining a new neutral element. Deduce that L · L can not be recognized by a finite group.

Notes for Chapter 8 The fact that the Stone dual of the Boolean algebra of regular languages over an alphabet is dual to the topological space underlying the free profinite monoid is well known but played a small role historically in the area. An early exception is the Pippenger (1997) article, see also Almeida (1995), and more recently Rhodes and Steinberg (2008), where the connection between profinite semigroups and Boolean rings with co-algebraic structure is exploited, and Gehrke et al. (2008, 2010), where the connection with modern topological duality as it is applied in logic was first developed. The point of these two last approaches is that it is not only the underlying space that can be studied via duality but the entire profinite semigroup or monoid structure. The duality-theoretic underpinnings of the work in Gehrke et al. (2008, 2010) were worked out in Gehrke (2016). For other key publications see Pin (2009, 2017), Branco and Pin (2009), and Gehrke et al. (2016). Starting with the definitions on page 280 and throughout this chapter, we fundamentally view the dual of an additional n-ary operation on a lattice or a Boolean algebra as an n + 1-ary relation on the dual space. Such a relation may also be seen as so-called co-algebraic structure. Co-algebraic structure is, by definition,

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

8.4 Open Multiplication

329

categorically dual to algebraic structure. For this reason, if we start, instead, from algebraic structure on the spatial side, then we get co-algebraic structure on the dual lattices or Boolean algebras. This formal duality point of view, for monoids, is exploited in Rhodes and Steinberg (2008, Section 8.4). What we show in Section 8.1 is that, if we restrict ourselves to surjective morphisms, then, in the setting of discrete duality, monoid structure on a set is dual to residuation structure on the dual Boolean algebras. This provides a kind of duality between algebras and algebras, rather than a duality between algebras and co-algebras, which is much more common. This duality between algebras and algebras can be lifted to non-surjective morphisms and to general Stone and Priestley duality as well, see Gehrke (2016). The definition of syntactic congruence (Definition 8.21) is classical in the theory of regular languages and can be traced back to the works Schützenberger (1955), Myhill (1957), and Nerode (1958). In Section 8.3, we encountered the notion of syntactic preorder, which generalizes syntactic congruences to an ordered setting, see Remark 8.46. This notion is already present in the work of Schützenberger (1955) and was pioneered by Pin (1995), long before the connection with Priestley duality was realized. We are grateful to Jean-Éric Pin and Howard Straubing for many enlightening discussions on the proof of Simon’s theorem discussed in Section 8.3. This proof will be the topic of a forthcoming joint paper of this book’s authors with Jérémie Marquès, who made significant contributions to the proof in this section, in particular Lemmas 8.64 and 8.65. Similar ideas, although not formulated using profinite methods, are present in Stern’s proof of Simon’s theorem (Stern, 1985). Many proofs of Simon’s theorem have been published since the original, see in particular also Almeida (1991), Higgins (1997), Henckell and Pin (2000), and Klíma (2011).

https://doi.org/10.1017/9781009349680.009 Published online by Cambridge University Press

Bibliography

Abramsky, S. 1987. Domain theory in logical form. In Proceedings of the Second Annual IEEE Symposium on Logic in Computer Science (LICS 1987). IEEE Computer Society Press, pp. 47–53. Abramsky, S. 1988. A cook’s Tour of the Finitary Non-well-founded Sets. Invited Lecture at BCTCS. Abramsky, S. 1990. The Lazy Lambda Calculus. Addison-Wesley Longman, pp. 65–116. Abramsky, S. 1991. Domain theory in logical form. Ann. Pur. Appl. Logic, 51(1), 1–77. Abramsky, S. and Jung, A. 1994. Domain Theory. Vol. 3. Oxford University Press, pp. 1–168. Abramsky, S. and Ong, C. H. L. 1993. Full abstraction in the lazy lambda calculus. Inf. Comput., 105(2), 159–267. Adámek, J., Herrlich, H., and Strecker, G. E. 1990. Abstract and Concrete Categories: The Joy of Cats. John Wiley and Sons. Adámek, J. and Rosický, J. 1994. Locally Presentable and Accessible Categories. LMS Lecture Note Series, vol. 189. Cambridge University Press. Almeida, J. 1991. Implicit operations on finite J -trivial semigroups and a conjecture of I. Simon. J. Pure. Appl. Algebra, 69(3), 205–218. Almeida, J. 1995. Finite Semigroups and Universal Algebra. World Scientific. Almeida, J. 2005. Profinite semigroups and applications. In Kudryavtsev, V. B., Rosenberg, I. G., and Goldstein, M. (eds.), Structural Theory of Automata, Semigroups, and Universal Algebra. Springer, pp. 1–45. Almeida, J. and Costa, A. 2009. Infinite-vertex free profinite semigroupoids and symbolic dynamics. J. Pure Appl. Algebra, 213, 605–631. Almeida, J. and Weil, P. 1998. Profinite categories and semidirect products. J. Pure Appl. Algebra, 123(1–3), 1–50. Almeida, J., Costa, A., Kyriakoglou, R., and Perrin, D. 2020. Profinite Semigroups and Symbolic Dynamics. Lecture Notes in Mathematics, vol. 2274. Springer International. Almeida, J., Goulet-Ouellet, H., and Klíma, O. 2023. What makes a Stone topological algebra profinite. Algebr. Univ., 84(6). Amadio, R. M. and Curien, P.-L. 1998. Domains and Lambda-Calculi. Cambridge Tracts in Theoretical Computer Science. Cambridge University Press. Awodey, S. 2010. Category Theory. 2nd ed. Oxford University Press.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

Bibliography

331

Balbes, R. and Dwinger, P. 1975. Distributive Lattices. University of Missouri Press. (Reprinted by Abstract Space Publishing in 2011). Banaschewski, B. 1996. Radical ideals and coherent frames. Comment. Math. Univ. Carolin., 37(2), 349–370. Barendregt, H. 2014. The Lambda Calculus: Its Syntax and Semantics. Studies in Logic and the Foundations of Mathematics, 2nd ed., vol. 103. North Holland. Barendregt, H. and Manzonetto, G. 2022. A Lambda Calculus Satellite. College Publications. Bekkali, M., Pouzet, M., and Zhani, D. 2007. Incidence structures and Stone–Priestley duality. Ann. Math. Artif. Intel., 49, 27–38. Bezhanishvili, G. 1998. Varieties of monadic Heyting algebras. Part I. Stud. Logica, 61, 367–402. Bezhanishvili, G. 1999. Varieties of monadic Heyting algebras. Part II. Duality theory. Stud. Logica, 62(1), 21–48. Bezhanishvili, G. 2000. Varieties of monadic Heyting algebras. Part III. Stud. Logica, 64(2), 215–256. Bezhanishvili, G., Bezhanishvili, N., Gabelaia, D., and Kurz, A. 2010. Bitopological duality for distributive lattices and Heyting algebras. Math. Struct. Comp. Sci., 20(3), 359–393. Bezhanishvili, G., Carai, L., and Morandi, P. J. 2022. The Vietoris functor and modal operators on rings of continuous functions. Ann. Pur. Appl. Logic, 173(1), 103029. Bezhanishvili, N. and Gehrke, M. 2011. Finitely generated free Heyting algebras via Birkhoff duality and coalgebra. Log. Meth. Comput. Sci., 7(2), 1–24. Bezhanishvili, N. and Kurz, A. 2007. Free modal algebras: A coalgebraic perspective. In Mossakowski, T., Montanari, U., and Haveraaen, M. (eds.), CALCO 2007. Lecture Notes in Computer Science, vol. 4624. Springer, pp. 143–157. Birkhoff, G. 1933. On the combination of subalgebras. Math. Proc. Cambridge, 29, 441–464. Birkhoff, G. 1967. Lattice Theory. 3rd ed. American Mathematical Society. Blackburn, P., de Rijke, M., and Venema, Y. 2001. Modal Logic. Cambridge Tracts in Theoretical Computer Science, vol. 53. Cambridge University Press. Boole, G. 1847. An Investigation of the Laws of Thought. Walton. Reprinted by Philisophical Library, New York, 1954. Borceux, F. 1994. Handbook of Categorical Algebra 1: Basic Category Theory. Cambridge University Press. Borceux, F. and Clementino, M. M. 2005. Topological semi-abelian algebras. Adv. Math., 190(2), 425–453. Borlido, C., Gehrke, M., Krebs, A., and Straubing, H. 2020. Difference hierarchies and duality with an application to formal languages. Topol. Appl., 273, 106975. Branco, M. J. J. and Pin, J.-E. 2009. Equations defining the polynomial closure of a lattice of regular languages. In Aceto, L., et al. (eds.), ICALP 2009, Part II. Lecture Notes in Computer Science, vol. 5556. Springer, pp. 115–126. Burris, S. and Sankappanavar, H. P. 2000. A Course in Universal Algebra: The Millennium Edition. www.math.uwaterloo.ca/snburris/htdocs/ualg.html Caramello, O. 2011. A Topos-Theoretic Approach to Stone-Type Dualities. arXiv:1103.3493. Caramello, O. 2017. Theories, Sites, Toposes. Oxford University Press.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

332

Bibliography

Celani, S., and Jansana, R. 1999. Priestley duality, a Sahlqvist theorem and a Goldblatt– Thomason theorem for positive modal logic. Log. J. IGPL, 7(6), 683–715. Celani, S. and Jansana, R. 2005. Bounded distributive lattices with strict implication. Math. Logic. Quart., 51(3), 219–246. Chagrov, A. and Zakharyaschev, M. 1997. Modal Logic. Oxford Logic Guides, vol. 35. Clarendon Press. Chikhladze, D., Clementino, M. M., and Hofmann, D. 2015. Representable (T, V)categories. Appl. Categor. Struct., 23(6), 829–858. Cignoli, R. L. O., D’Ottaviano, I. M. L., and Mundici, D. 2000. Algebraic Foundations of Many-Valued Reasoning. Trends in Logic – Studia Logica Library, vol. 7. Kluwer Academic Publishers. Clark, D. M. and Davey, B. A. 1998. Natural Dualities for the Working Algebraist. Cambridge Studies in Advanced Mathematics. Cambridge University Press. Conradie, W., Ghilardi, S., and Palmigiano, A. 2014. Unified Correspondence. Outstanding Contributions to Logic. Springer International, pp. 933–975. Conradie, W. and Palmigiano, A. 2020. Constructive canonicity of inductive inequalities. Log. Meth. Comput. Sci., 16(3), 8:1–8:39. https://lmcs.episciences.org/6694/pdf Cornish, W. H. 1975. On H. Priestley’s dual of the category of bounded distributive lattices. Matematički Vesnik, 12(27)(60), 329–332. Coumans, D. C. S. and Gool, S. J. van. 2012. On generalizing free algebras for a functor. J. Logic. Comput., 23(3), 645–672. Davey, B. A. and Priestley, H. A. 2002. Introduction to Lattices and Order. 2nd ed. Cambridge University Press. Dickmann, M., Schwartz, N., and Tressl, M. 2019. Spectral Spaces. New Mathematical Monographs. Cambridge University Press. Dimov, G., Ivanova-Dimova, E., and Tholen, W. 2022. Categorical extension of dualities: From Stone to de Vries and beyond. Appl. Categ. Struct., 30, 287–329. Dunn, J. M. and Hardegree, G. 2001. Algebraic Methods in Philosophical Logic. Oxford University Press. Dzamonja, M. 2008. Representation theorems for connected compact Hausdorff spaces. Sarajevo J. Math., 4(16), 7–21. Eilenberg, S. 1974. Automata, Languages, and Machines. Vol. A. Academic Press. Pure and Applied Mathematics, Vol. 59A. Eilenberg, S. 1976. Automata, Languages, and Machines. Vol. B. Academic Press. Pure and Applied Mathematics, Vol. 59B. Engelking, R. 1989. General Topology. Sigma Series in Pure Mathematics, vol. 6. Heldermenn Verlag. Erné, M. 2009. Quasicoherent domains and hyperspectral spaces: A missing link in Stone– Priestley duality. Abstract of a talk given at Topology, Algebra, and Categories in Logic (TACL). Esakia, L. 1974. Topological Kripke models. Dokl. Akad. Nauk SSSR, 15(1), 147–151. Esakia, L. 2019. Heyting Algebras: Duality Theory. Trends in Logic, vol. 50. Springer. Translation of the Russian 1985 original. Fleisher, I. 2000. Priestley’s duality from Stone’s. Adv. Appl. Math., 25(3), 233–238. Forssell, H. 2008. First-Order Logical Duality. PhD thesis, Carnegie Mellon University. Fussner, W. and Palmigiano, A. 2019. Residuation algebras with functional duals. Algebr. Univ., 80(4), 1–10.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

Bibliography

333

Galatos, N., Jipsen, P., Kowalski, T., and Ono, H. 2007. Residuated Lattices: An Algebraic Glimpse at Substructural Logics. Studies in Logic and the Foundations of Mathematics. Elsevier Science. Gastin, P., Manuel, A., and Govind, R. 2021. Reversible regular languages: Logical and algebraic characterisations. Fund. Inform., 180(4), 333–350. Gehrke, M. 2014. Canonical extensions, Esakia spaces, and universal models. In Bezhanishvili, G. (ed.), Leo Esakia on Duality in Modal and Intuitionistic Logics. Springer, pp. 9–41. Gehrke, M. 2016. Stone duality, topological algebra, and recognition. J. Pure Appl. Algebra, 220(7), 2711–2747. Gehrke, M. 2018. Canonical extensions: An algebraic approach to Stone duality. Algebr. Univ., 79(3). Gehrke, M. and Gool, S. J. van. 2014. Distributive envelopes and topological duality for lattices via canonical extensions. Order, 31(3), 435–461. Gehrke, M., Grigorieff, S., and Pin, J.-E. 2008. Duality and equational theory of regular languages. In Aceto, L. et al. (eds.), ICALP 2008, Part II. Lecture Notes in Computer Science, vol. 5126. Springer, pp. 246–257. Gehrke, M., Grigorieff, S., and Pin, J.-E. 2010. A topological approach to recognition. In Abramsky, S. et al. (eds.), Automata, Languages and Programming (ICALP 2010). Lecture Notes in Computer Science, vol. 6199, no. 2. Springer, pp. 151–162. Gehrke, M. and Jónsson, B. 1994. Bounded distributive lattices with operators. Math. Japon., 40(2), 207–215. Gehrke, M. and Jónsson, B. 2004. Bounded distributive lattice expansions. Math. Scand., 94, 13–45. Gehrke, M., Krebs, A., and Pin, J.-E. 2016. Ultrafilters on words for a fragment of logic. Theor. Comput. Sci., 610( Part A), 37–58. Gehrke, M., Nagahashi, H., and Venema, Y. 2005. A Sahlqvist theorem for distributive modal logic. Ann. Pur. Appl. Logic, 131(1–3), 65–102. Gehrke, M. and Priestley, H. A. 2006. Duality for distributive quasioperator algebras via their canonical extensions. Stud. Logica, 86(1), 31–68. Gehrke, M. and Priestley, H. A. 2007. Canonical extensions of double quasioperator algebras: An algebraic perspective on duality for certain algebras with binary operations. J. Pure Appl. Algebra, 209(1), 269–290. Gehrke, M., Jakl, T., and Reggio, L. 2023. A cook’s tour of duality in logic: From quantifiers, through Vietoris, to Measures. In Palmigiano, A. and Sadrzadeh, M. (eds.), Samson Abramsky on Logic and Structure in Computer Science and Beyond. Springer, pp. 129–158. Ghilardi, S. 1992. Free Heyting algebras as bi-Heyting algebras. C. R. Math. Rep. Acad. Sci. Canada, 14(6), 240–244. Ghilardi, S. 1995. An algebraic theory of normal forms. Ann. Pur. Appl. Logic, 71, 189–245. Ghilardi, S. 2010. Continuity, freeness, and filtrations. J. Appl. Non-Class. Log., 20(3), 193–217. Gierz, G., Hofmann, K. H., Keimel, K., et al. 1980. A Compendium of Continuous Lattices. Springer-Verlag. Gierz, G., Hofmann, K. H., Keimel, K., et al. 2003. Continuous Lattices and Domains. Encyclopedia of Mathematics and its Applications. Cambridge University Press. Givant, S. 2014. Duality Theories for Boolean Algebras with Operators. Springer. Givant, S. and Halmos, P. 2008. Introduction to Boolean Algebras. Undergraduate Texts in Mathematics. Springer-Verlag.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

334

Bibliography

Goldblatt, R. 1989. Varieties of complex algebras. Ann. Pur. Appl. Logic, 44, 173–242. Goldblatt, R., Hodkinson, I., and Venema, Y. 2003. On canonical modal logics that are not elementarily determined. Log. Anal., 181, 77–101. Gool, S. J. van. 2014. On Sheaves and Duality. PhD thesis, Raboud University Nijmegen. Gool, S. J. van, and Marquès, J. 2024. On duality and model theory for polyadic spaces. Ann. Pur. Appl. Logic, 175 (2), 103388. Goubault-Larrecq, J. 2013. Non-Hausdorff Topology and Domain Theory. Cambridge University Press. Goubault-Larrecq, J. 2019. A probabilistic and non-deterministic call-by-push-value language. In 34th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS), IEEE, pp. 1–13. Grätzer, G. 2003. General Lattice Theory. 2nd ed. Birkhäuser. Grätzer, G. 2008. Universal Algebra. 2nd ed. Springer. Grätzer, G. 2011. Lattice Theory: Foundation. Birkhäuser. Gunter, C. 1985. Profinite Solutions for Recursive Domain Equations. PhD thesis, University of Wisconsin. Hart, K. P. and van Mill, J. 2022. Problems on βN. arXiv:2205.11204. Hausdorff, F. 1914. Grundzüge der Mengenlehre. Veit & Comp. Reprinted by Chelsea Publishing Company in 1949. Haviar, M. 2019. On selected developments in the theory of natural dualities. Acta Univ. M. Belii Ser. Math., 27, 59–78. Henckell, K. and Pin, J.-E. 2000. Ordered monoids and J -trivial monoids. In Birget, J-C., Margolis, S., Meakin, J., and Sapir, M. (eds.), Algorithmic Problems in Groups and Semigroups (Lincoln, NE, 1998). Trends Math. Birkhäuser, pp. 121–137. Higgins, P. M. 1997. A proof of Simon’s theorem on piecewise testable languages. Theor. Comput. Sci., 178(1–2), 257–264. Higman, G. 1952. Ordering by divisibility in abstract algebras. Proc. London Math. Soc., 3(1), 326–336. Ho, W. K., Goubault-Larrecq, J., Jung, A., and Xi, X. 2018. The Ho–Zhao problem. Log. Meth. Comput. Sci., 14(1). Hodkinson, I. and Venema, Y. 2005. Canonical varieties with no canonical axiomatisation. T. Am. Math. Soc., 357, 4579–4605. Hoffmann, R.-E. 1981. Continuous posets, prime spectra of completely distributive complete lattices, and Hausdorff compactification. In Banaschewski, B. and Hoffmann, R.-E. (eds.), Continuous Lattices, Proceedings Bremen 1979. Lecture Notes in Mathematics, vol. 871. Springer Verlag, pp. 159–208. Hofmann, K. H. and Lawson, J. D. 1978. The spectral theory of distributive continuous lattices. T. Am. Math. Soc., 246, 285–310. Hofmann, K. H., Mislove, M., and Stralka, A. 1974. The Pontryagin Duality of Compact 0-Dimensional Semilattices and its Applications. Springer. Hofmann, D. and Nora, P. 2015. Dualities for modal algebras from the point of view of triples. Algebr. Univ., 73, 297–320. Hofmann, D. and Nora, P. 2018. Enriched Stone-type dualities. Adv. Math., 330, 307–360. Hofmann, D. and Nora, P. 2023. Duality theory for enriched Priestley spaces. J. Pure Appl. Algebra, 227(3), 107231. Hofmann, D., Seal, G. J., and Tholen, W. 2014. Monoidal Topology: A Categorical Approach to Order, Metric, and Topology. Cambridge University Press.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

Bibliography

335

Hofmann, D. and Stubbe, I. 2011. Towards Stone duality for topological theories. Topol. Appl., 158(7), 913–925. Howard, P. and Rubin, J. E. 1998. Consequences of the Axiom of Choice. American Mathematical Society. Jech, T. J. 1973. The Axiom of Choice. North Holland. Jia, X., Jung, A., and Q., Li. 2016. A note on coherence of dcpos. Topology and its Applications, 209, 235–238. Jipsen, P. 2009. Generalizations of Boolean products for lattice-ordered algebras. Ann. Pur. Appl., 161, 228–234. Johnstone, P. T. 1981. Scott is not always sober. In Banaschewski, B. and Hoffmann, R.-E. (eds.), Continuous Lattices. Lecture Notes in Mathematics, vol. 871. Springer, pp. 282–283. Johnstone, P. T. 1986. Stone Spaces. Cambridge University Press. Reprint of the 1982 ed. Jónsson, B. and Tarski, A. 1951. Boolean algebras with operators. I. Am. J. Math., 73(4), 891–939. Jónsson, B. and Tarski, A. 1952. Boolean algebras with operators. II. Am. J. Math., 74(1), 127–162. Jung, A. 1989. Cartesian Closed Categories of Domains. Tech. Rept. Amsterdam. Jung, A. 1990. The classification of continuous domains. In Proceedings. Fifth Annual IEEE Symposium on Logic in Computer Science. IEEE Computer Society Press, pp. 35–40. Jung, A. 2004. Stably compact spaces and the probabilistic powerspace construction. Electron. Notes Theor. Comput. Sci., 87, 15. Jung, A. and Moshier, M. A. 2006. On the Bitopological Nature of Stone Duality. Tech. Rept. CSR-06-13. University of Birmingham. Jung, A. and Sünderhauf, P. 1996. On the duality of compact vs. open. In Andima, S., Flagg, R.C., Itzkowitz, G., et al. (eds.), Papers on General Topology and Applications: Eleventh Summer Conference at the University of Southern Maine. vol. 806. New York Academy of Sciences, pp. 214–230. Jung, A. and Tix, R. 1998. The troublesome probabilistic powerdomain. Electron. Notes Theor. Comput. Sci., 13, 70–91. Kleene, S. C. 1956. Representation of events in nerve nets and finite automata. In Shannon, C. E. and McCarthy, J. (eds.), Automata Studies. Annals of Mathematics Studies, vol. 34. Princeton University Press, pp. 3–41. Klíma, O. 2011. Piecewise testable languages via combinatorics on words. Discrete Math., 311(20), 2124–2127. Koppelberg, S., Monk, J. D., and Bonnet, R. 1989. Handbook of Boolean Algebras. Vol. 1. North-Holland. Kupke, C., Kurz, A., and Venema, Y. 2004. Stone coalgebras. Theor. Comput. Sci., 327(1–2), 109–134. Lambek, J. and Scott, P. J. 1986. Introduction to Higher Order Categorical Logic. Cambridge University Press. Larsen, K. G. and Winskel, G. 1991. Using information systems to solve recursive domain equations. Inform. Comput., 91(2), 232–258. Lawson, J. D. 1979. The duality of continuous posets. Houston J. Math., 5, 357–394. Lawson, J. D. 2011. Stably compact spaces. Math. Struct. Comp. Sci., 21, 125–169.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

336

Bibliography

Leinster, T. 2014. Basic Category Theory. Cambridge University Press. Lurie, J. 2019. Ultracategories. Preprint. https://people.math.harvard.edu/∼lurie/papers/ Conceptual.pdf Mac Lane, S. 1971. Categories for the Working Mathematician. Graduate Texts in Mathematics, vol. 5. Springer-Verlag. Mac Lane, S. and Moerdijk, I. 1992. Sheaves in Geometry and Logic: A First Introduction to Topos Theory. Springer. MacNeille, H. M. 1937. Partially ordered sets. Trans. Amer. Math. Soc., 42, 416–460. Makkai, M. 1987. Stone duality for first-order logic. Adv. Math., 65, 97–170. Makkai, M. 1993. Duality and definability in first order logic. Mem. Amer. Math. Soc., 105(503), 97–170. Makkai, M., and Reyes, G. 1977. First Order Categorical Logic. Lecture Notes in Mathematics, no. 611. Springer. Manzonetto, G. and Salibra, A. 2008. Applying universal algebra to lambda calculus. J. Logic Comput., 20(4), 877–915. Marquès, J. 2021. Polyadic spaces and profinite monoids. In Fahrenberg, U., Gehrke, M., Santocanale, L., and Winter, M. (eds.), Relational and Algebraic Methods in Computer Science: 19th International Conference, RAMiCS 2021, Marseille, France, November 2–5, 2021, Proceedings. Springer International Publishing, pp. 292–308. Marquis, J.-P. and Reyes, G. 2011. The history of categorical logic: 1963–1977. In Gabbay, D., Kanamori, A., and Woods, J. (eds.), Handbook of the History of Logic. Elsevier, pp. 1–116. McKinsey, J. C. C. and Tarski, A. 1948. Some theorems about the sentential calculi of Lewis and Heyting. J. Symbolic Logic, 13(1), 1–15. Moshier, M. A. and Jipsen, P. 2014a. Topological duality and lattice expansions, I: A topological construction of canonical extensions. Algebr. Univ., 71, 109–126. Moshier, M. A. and Jipsen, P. 2014b. Topological duality and lattice expansions, II: Lattice expansions with quasioperators. Algebr. Univ., 71, 221–234. Mundici, D. 2011. Advanced Łukasiewicz Calculus and MV-Algebras. Trends in Logic – Studia Logica Library. Vol. 35. Springer. Murthy, C. R and Russell, J. R. 1990. A constructive proof of Higman’s lemma. In Proceedings of the Fifth Annual IEEE Symposium on Logic in Computer Science. IEEE, pp. 257–267. Myhill, J. 1957. Finite Automata and the Representation of Events. Tech. Rept. WADD. Nachbin, L. 1947. Une propriété caractéristique des algèbres booléiennes. Port. Math., 6, 115–118. Nachbin, L. 1964. Topology and Order. van Nostrand. Nerode, A. 1958. Linear automaton transformations. P. Am. Math. Soc., 9(4), 541–544. Nerode, A. 1959. Some Stone spaces and recursion theory. Duke Math. J., 26, 397–406. Palmigiano, A. 2004. A coalgebraic view on positive modal logic. Theor. Comput. Sci., 327(1), 175–195. Picado, J. 1994. Join-continuous frames, Priestley’s duality and biframes. Appl. Categor. Struct., 2, 297–313. Picado, J. and Pultr, A. 2012. Frames and Locales. Birkhäuser. Pin, J.-E. 1995. A variety theorem without complementation. Izv. Math. (Izvestija vuzov. Matematika), 39, 80–90. Pin, J.-E. 2009. Profinite methods in automata theory. In Albers, S. and Marion, J.-Y. (eds.), 26th International Symposium on Theoretical Aspects of Computer Science (STACS 2009), vol. 5556, Schloss Dagstuhl, pp. 31–50. Pin, J.-E. 2011. Theme and variations on the concatenation product. In Winkler, F. (ed.), CAI. Springer, pp. 44–64.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

Bibliography

337

Pin, J.-E. 2017. Dual space of a lattice as the completion of a Pervin space. In Höfner, P., Pous, D., and Struth, G. (eds.), RAMICS. Lecture Notes in Computer Science, vol. 10226. Springer, pp. 151–162. Pin, J.-E. (ed.). 2021a. Handbook of Automata Theory. Volume I. Theoretical Foundations. European Mathematical Society (EMS). Pin, J.-E. (ed.). 2021b. Handbook of Auomata Theory. Volume II. Automata in Mathematics and Selected Applications. European Mathematical Society (EMS). Pin, J.-E. 2022. Mathematical Foundations of Automata Theory. Lecture notes. https://www.irif.fr/∼jep/PDF/MPRI/MPRI.pdf Pippenger, N. 1997. Regular languages and Stone duality. Theor. Comput. Syst., 30(2), 121–134. Plotkin, G. D. 1976. A powerdomain construction. SIAM J. Comput., 5(3), 452–487. Pouzet, M. 2018. A topological interpretation of de Jongh-Parikh theorem. https://www2 .cms.math.ca/Events/winter18/abs/pdf/sdr-mp.pdf Priestley, H. A. 1970. Representation of distributive lattices by means of ordered Stone spaces. Bull. London Math. Soc., 2, 186–190. Raney, G. N. 1953. A subdirect-union representation for completely distributive complete lattices. P. Am. Math. Soc., 4(4), 518–522. Reiterman, J. 1982. The Birkhoff theorem for finite algebras. Algebr. Univ., 14(1), 1–10. Rhodes, J., Schilling, A., and Silva, P. V. 2022. Holonomy theorem for finite semigroups. Int. J. Algebr. Comput., 32(03), 443–460. Rhodes, J. and Steinberg, B. 2008. The Q-Theory of Finite Semigroups. Springer. Ribes, L. and Zalesskii, P. 2010. Profinite Groups. Springer. Sakarovitch, J. 2009. Elements of Automata Theory. Cambridge University Press. Santocanale, L. 2020. The involutive quantaloid of completely distributive lattices. In Relational and Algebraic Methods in Computer Science: 18th International Conference, RAMiCS, Palaiseau, France, October 26–29. Proceedings. Springer, pp. 286–301. Santocanale, L., and Wehrung, F. 2014. Lattices of regular closed subsets of closure spaces. Int. J. Algebr. Comput., 24(07), 969–1030. Schmid, J. 2002. Quasiorders and sublattices of distributive lattices. Order, 19(1), 11–34. Schützenberger, M. P. 1955. Une théorie algébrique du codage. Séminaire Dubreil. Algèbre et Théorie des Nombres, 9, 1–24. Schützenberger, M. P. 1965. On finite monoids having only trivial subgroups. Inform. Cont., 8(2), 190–194. Scott, D. S. 1972. Continuous Lattices. In Lawvere, F. W. (ed.), Toposes, Algebraic Geometry, and Logic. Lecture Notes in Computer Science, vol. 274. Springer, pp. 97–136. Scott, D. S. 1980. Relating theories of the lambda calculus. In Hindley, J. R. and Seldin, J. P. (eds.), To H. B. Curry: Essays in Combinatory Logic, Lambda Calculus and Formalism. Academic Press, pp. 403–450. Scott, D. S. and Strachey, C. 1971. Towards a mathematical semantics for computer languages. In Fox, J. (ed.), Proceedings of the Symposium on Computers and Automata. Microwave Research Institute Symposia Series. Polytechnic Institute of Brooklyn Press. Also Tech. Mon. PRG-6, Oxford University Computing Laboratory, pp. 19–46. Simon, I. 1975. Piecewise testable events. In Brackage, H. (ed.), Proc. 2nd GI Conf. Lecture Notes in Comp. Sci., vol. 33. Springer Verlag, pp. 214–222.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

338

Bibliography

Smyth, M. B. 1983a. The largest Cartesian closed category of domains. Theor. Comput. Sci., 27, 109–119. Smyth, M. B. 1983b. Power domains and predicate transformers: A topological view. In Diaz, Josep (ed.), Automata, Languages and Programming. Springer, pp. 662–675. Smyth, M. B. 1992. Stable compactification I. J. London Math. Soc., 45, 321–340. Smyth, M. B. and Plotkin, G. D. 1982. The C-category-theoretic solution of recursive domain equations. SIAM Journal on Computing, 11(4), 761–783. Steinberg, B. 2013. Topological dynamics and recognition of languages. arXiv:1306.1468. Stern, J. 1985. Characterizations of some classes of regular events. Theor. Comput. Sci., 35(1), 17–42. Stone, M. H. 1934. Boolean algebras and their application to topology. Proc. Natl. Acad. Sci. USA, 20(3), 197–202. Stone, M. H. 1935. Subsumption of the theory of Boolean algebras under the theory of rings. Proc. Natl. Acad. Sci. USA, 21(2), 103–105. Stone, M. H. 1936. The theory of representation for Boolean algebras. T. Am. Math. Soc., 40(1), 37–111. Stone, M. H. 1937a. Applications of the theory of Boolean rings to general topology. T. Am. Math. Soc., 41, 375–481. Stone, M. H. 1937b. Topological representations of distributive lattices and Brouwerian logics. Čas. Mat. Fys., 67, 1–25. Stone, M. H. 1938. The representation of Boolean algebras. B. Am. Math. Soc., 44(12), 807–816. Stralka, A. 1980. A partially ordered space which is not a Priestley space. Semigroup Forum, 20, 293–297. Straubing, H., and Thérien, D. 1988. Partially ordered finite monoids and a theorem of I. Simon. J. Algebra, 119(2), 393–399. Venema, Y. 2007. Algebras and coalgebras. In Studies in Logic and Practical Reasoning, vol. 3. Elsevier, pp. 331–426. Vickers, S. 1989. Topology via Logic. Cambridge University Press. Wechler, W. 1992. Universal Algebra for Computer Scientists. EATCS Monographs on Theoretical Computer Science, vol. 25. Springer-Verlag. Wehrung, F. 2008. Poset representations of distributive semilattices. Int. J. Algebr. Comput., 18(2), 321–356. Weil, P. 2002. Profinite methods in semigroup theory. Int. J. Algebr. Comput., 12, 137–178. Zhang, G.-Q. 1989. Logics of domains. Tech. Rept. UCAM-CL-TR-185. University of Cambridge, Computer Laboratory. Zhang, G.-Q. 1991. Logic of Domains. Progress in Theoretical Computer Science. Birkhauser.

https://doi.org/10.1017/9781009349680.010 Published online by Cambridge University Press

Notation

p −< q Pop n 2 g◦ f f [S] ker f f¯ L− a+b D(P) U (P) ↓S ↑S ↑p ↓p Uc J (L) M(L)

Published online by Cambridge University Press

q covers p, q is an immediate successor of p, 2 the opposite of a poset P, 3 the n-element, 3 the two-element, 3 composition of the functions f and g, g after f , 3 forward image under a function f of a subset S of the domain, 10 kernel of a homomorphism f , 11 unique extension of a function f given by a universal property determined by the context, 14 Boolean envelope of a distributive lattice L, 15 symmetric difference of two Boolean algebra elements a and b, 17 the complete lattice of down-sets of a poset P, 18 the complete lattice of up-sets of a poset P, 18 the down-set generated by a subset S of a poset P, 19 the up-set generated by a subset S of a poset P, 19 the principal up-set generated by an element p of a poset P, 19 the principal down-set generated by an element p of a poset P, 19 the complement of a subset U, 19 the poset of (finitely) join-irreducible elements of a lattice L, 19 the poset of (finitely) meet-irreducible elements of a lattice L, 19

340

Notation

aˆ f −1 (S) Ω(X ) C(X ) Clp(X ) int(S) S◦ cl(S) S ≤τ K S(X ) ι↑ (P) σ(P) α(P) τ↑ τ↓ X↑ X↓ τ∂ τp P⊕Q Filt(L) PrFilt(L) Idl(L) PrIdl(L) aˆ

Published online by Cambridge University Press

representation of an element a of a finite distributive lattice as a down-set of its dual space, 20 inverse image of a subset S of the codomain under a function f , 21 the collection of open subsets of a topological space X, 29 the collection of closed subsets of a topological space X, 29 the collection of clopen subsets of a topological space X, 29 interior of a subset S of a topological space, 30 interior of a subset S of a topological space, 30 closure of a subset S of a topological space, 31 closure of a subset S of a topological space, 31 specialization order associated with a topology τ, 37 the collection of compact-saturated subsets of a topological space X, 37 the upper topology on a poset P, 39 the Scott topology on a poset P, 39 the Alexandrov topology on a poset P, 40 the open up-set topology associated with an ordered space (X, τ, ≤), 43 the open down-set topology associated with an ordered space (X, τ, ≤), 43 the topological space with open up-set topology associated with an ordered space X, 43 the topological space with open down-set topology associated with an ordered space X, 43 the co-compact dual of a topology τ, 45 the patch topology obtained from a topology τ, 45 ordered sum of two posets P and Q, with P below Q, 51 the poset of filters of a lattice L, 53 the poset of prime filters of a lattice L, 53 the poset of ideals of a lattice L, 56 the poset of prime ideals of a lattice L, 56 representation of an element a of a distributive lattice as a down-set of its dual space, 58

Notation

Sfilt Sidl XL Fx Ix hx βX ∗X

FDL (V ) a≈b x |= a ≈ b

a ≈ b R ϑ(x) ϑ(S) R a |= x  y

a A (x, y)I

Published online by Cambridge University Press

filter generated by a subset S of a lattice L, 60 ideal generated by a subset S of a lattice L, 61 the (neutral) dual space of a distributive lattice L, 65 prime filter corresponding to a point x of a (neutral) dual space, 65 prime ideal corresponding to a point x of a (neutral) dual space, 65 homomorphism into 2 corresponding to a point x of a (neutral) dual space, 65 Stone–Čech compactification of a discrete space X, 80 remainder of the Stone–Čech compactification of the (discrete) space X, 81 free distributive lattice over a set V , 88 pair of lattice elements, seen as equation on the dual space, 97 satisfaction relation between an element x of a dual space and a pair (a, b) of elements of a lattice, viewed as an equation, 97 space associated with an equation between lattice elements a and b, 97 space associated with a binary relation R on a lattice, 97 congruence on a lattice associated with an element x of its dual space, 97 congruence on a lattice associated with a subset S of its dual space, 97 the lattice congruence generated by R, for R a binary relation on a lattice L, 98 satisfaction relation between an element a of a lattice and a pair (x, y) of elements of the dual space viewed as an inequation, 102 preorder on the dual space associated with an element a of a lattice, 103 preorder on the dual space associated with a subset A of a lattice, 103 sublattice associated with a pair of elements of the dual space viewed as an inequation, 103

341

342

Notation

RI a |= x ≡ y

A |= x ≡ y

≡a ≡A (x, y)E

RE

≡ R·S R[U] ∃R [U] ∀R [U]

K |= ϕ R[U, _, _] R[U, V, _] ⇒R

Published online by Cambridge University Press

sublattice associated with a binary relation on the dual space viewed as a set of inequations, 103 satisfaction relation between an element a of a Boolean algebra and a pair (x, y) of elements of the dual space, viewed as an equation, 105 satisfaction relation between a subset A of a Boolean algebra and a pair (x, y) of elements of the dual space, viewed as an equation, 105 equivalence relation on the dual space associated with an element a of a Boolean algebra, 105 equivalence relation on the dual space associated with a subset A of a Boolean algebra, 105 Boolean subalgebra associated with a pair of elements of the dual space viewed as an equation, 105 Boolean subalgebra associated with a binary relation R on the dual space viewed as a set of equations, 105 Boolean subalgebra of P (S) associated with an equivalence relation ≡ on a set S, 109 relational composition of two relations R and S, with R before S, 112 direct image under a binary relation R of a subset U of the domain, 113 direct image under a binary relation R of a subset U of the domain, 113 relational universal image under a binary relation R of a subset U of the domain, 113 derivability in normal modal logic K, 122 forcing relation on Kripke models, 123 truth set of a modal formula ϕ in a Kripke model, 123 direct image of a subset U under a ternary relation R, 133 direct image of subsets U and V under a ternary relation R, 133 operator of implication type associated with a ternary relation R, 134

Notation

u·v u\v u/v a→b

ob C mor C HomC ( A, B) Cop ϕ: F ⇒ G A∗ K O(X ) N (x) CompPrFilt(L) Pt(L) a˜ [X, Y ] K ⇒U

K(X ) x)y  

y y UB(M) MUB(M)

Published online by Cambridge University Press

binary operator associated to a ternary relation R, equal to R[u, v, _], 138 operator of implication type associated to a ternary relation R, left residual of ·, 138 operator of implication type associated to a ternary relation R, right residual of ·, 138 relative pseudocomplement of a with respect to b, also known as the Heyting implication or Heyting arrow, 139 objects of a category C, 155 morphisms of a category C, 155 morphisms from A to B in a category C, 155 the opposite of a category, 159 a natural transformation ϕ from a functor F to a functor G, 162 free monoid over a set A, 168 the set of compact-open subsets of a topological space X, 198 the neighborhood filter of a point x in a topological space, 201 the set of completely prime filters of a frame L, 202 the space of points of a frame L, 203 the subset of the filter space defined by an element a, 213 the space of continuous functions from X to Y , 217 the set of functions f for which the direct image of K is contained in U; a generic open set in the subbase for the topology on [X, Y ], 217 the set of compact elements of a dcpo X, 235 the way-below relation between elements x and y of a dcpo, 235 the set of elements way below y, 235 the set of elements way above y, 235 the set of upper bounds of a subset M of a poset, 253 the set of minimal upper bounds of a subset M of a poset, 253

343

344

Notation

1M δx L(A) L(I, F)

R (M) L(M) B(M) x −1 L Ly −1 Q(S) L(L) B(L) ≡s Ms Reg A fˆ

h¯

Aˆ∗ ClpCon(M) ∗

L |= x  y

Published online by Cambridge University Press

neutral element of a monoid M, 275 reachability relation of a machine with transition relation δ when reading a word x, 276 language recognized by an automaton A, 277 language recognized by a finite state machine M when taking I as the set of initial states and F as the set of final states, 277 set of languages recognizable by a finite state machine M, 277 lattice of languages generated by R (M), 277 Boolean algebra of languages generated by R (M), 277 left quotient of a language L by the word x, 278 right quotient of a language L by the word y, 278 closure under quotienting of a subset S of the residuated Boolean algebra dual to a monoid, 281 the residuation ideal generated by a language L, 281 the Boolean residuation ideal generated by a language L, 281 syntactic congruence of a subset s of a monoid M, 288 syntactic monoid of a subset s of a monoid M, 288 Boolean algebra of regular languages over a finite alphabet A, 289 the unique continuous homomorphic extension to the free profinite monoid of a function f from generators to a finite monoid, 295 the unique continuous extension to the free profinite monoid of a homomorphism h from the free monoid to a finite monoid, 295 the free profinite monoid over a finite set A, 300 the set of clopen congruences on a topological monoid M, 300 a regular language L satisfies a profinite monoid inequation given by the pair (x, y) of elements of the free profinite monoid, 312

Notation ∗

x  y ∗

L |= x = y

+ PT+ PT

Published online by Cambridge University Press

the set of regular languages L which satisfy a profinite monoid inequation given by the pair (x, y) of elements of the free profinite monoid, 312 a regular language L satisfies a profinite monoid equation given by the pair (x, y) of elements of the free profinite monoid, 312 the subword partial order, 314 the collection of positively piecewise testable languages (in a fixed finite alphabet), 314 the collection of piecewise testable languages (in a fixed finite alphabet), 314

345

Index

absorption laws, 9 accepting state, 275 accepts, 275 action left, 283 right, 283 two-sided, 283 adjoint functor theorem for complete lattices, 17 adjoint pair, 119 dual relation, 121 adjunction between categories, 166 between preorders, 6 admissible valuation, 123 Alexandrov topology, 40 algebra for an endofunctor, 194 algebraic dcpo, 235 algebraic domain, 208, 228, 235 alternating chain, 149 alternation height, 149 anti-chain, 1 anti-isomorphism, 3 antitone, 3 aperiodic, 318 approximable, 245 associative, 9 atom, 23 atomic, 24, 209 atomless, 95 automaton, 275 axiom of choice, 33 back-and-forth morphisms, 127 base, 31 biaction, 283 biaction morphism, 284 bifinite, 260–262 2/3 bifinite, 255 bilinear maps, 132 binary topological algebra, 302 Birkhoff duality, 23, 24

Boolean algebra, 14 dual space of, 79 Boolean compatible relation, 117 Boolean envelope, 14, 24, 76, 94, 148, 168 Boolean residuation algebra, 294 Boolean space, 79 Booleanization, 14 bottom, 4 bounded morphism, 127, 136, 144 butterfly poset, 5 canonical, 128, 153 canonical extension, 126, 168 canonical model, 126 Cantor space, 49, 83, 92 ordered, 92, 212 Cartesian closed category, 227 Cartesian product, 10 category, 155 co-complete, 185 concrete, 156 large, 155 locally small, 155 monoid as a, 156 preorder as a, 156 skeleton, 157, 171 small, 155 thin, 156, 157 Cayley biaction, 284 Cayley representation, 284 chain, 1 choice functions, 238 clopen, 29 closed mapping, 30 closed sets, 29 closed under quotienting, 278, 280 closure, 31 closure under quotienting, 280 co-compact dual, 45 co-cone, 179 co-finite, 16, 61

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

Index co-Heyting algebra, 139, 147 co-restriction, 34 co-span, 183 co-unit, 166 coequalizer, 174 is an epimorphism, 186 coherent, 44, 248 coherent spaces, see under spectral space colimit, 179 commutative, 9 compact element, 235 compact ordered space, 42 compact space, 32 compact-open topology, 217 compact-saturated, 38 compactification, 80 comparable, 1 compatible downward Priestley, 118 equivalence relation, 105 monoid, 312 preorder, 101 relation of implication type, 133 relation of implication type on a Boolean space, 302 relation on a Priestley space, 118 relation on a Boolean space, 117 relation on a Priestley space, 114 upward order, 113 upward Priestley, 114 complement, 19 complete automaton, 276 complete Heyting algebra, 12 complete implication-type operator, 130 complete lattice, 9 complete operator of diamond type, 130, 137 complete sublattice, 11 completely distributive, 238 completely join irreducible, 244 completely join prime, 26, 240 elements generate, 26 completely prime filter, 202 concatenation, 3 concrete categories, 161 cone, 179 congruence clopen, 295 on a lattice, 11 on a monoid, 282 conjunctively closed, 262 consistent intuitionistic formula, 145 constant diagram, 187 continuous, 29 proper, 50 continuous dcpo, 235 contravariant adjunction, 6 converse relation, 119 and adjunction, 119, 121

convex set, 19 coproduct in categories, 177 special case of colimit, 179 CTOD, 63 dcpo, 14 space of a, 233 spectral, 228 dense, 31 deterministic automaton, 276 diagram constant, 187 diamond poset, 2 dictionary order, 4 difference chain, 149 direct image, 11 directed, 13 directed join, 13 directedly complete poset, 14 discrete category, 179 discrete duality, 23, 24, 26 discrete monoid, 295 discrete topology, 30, 40 distributivity, 12 failure of, 13 divisibility, 54 domain, 228, 235 algebraic, 235 bifinite, 229 spectral, 228 down-directed, 13 down-set, 18 generated by a subset, 19 principal, 19 down-set lattice, 26 dual Alexandrov topology, 40 dual categories, 164 dual distributive lattice, 21 dual isomorphism, 170 dual poset, 21 dual relation, 134 duality Birkhoff, 23 for finite Boolean algebras, 24 for finite distributive lattices, 23 Priestley, 73 Stone, 79, 196 dually equivalent categories, 164 dually isomorphic, 170 embedding projection pair, 257 enough join irreducibles, 19 epimorphism, 174 regular, 174 equalizer, 175, 179 is a monomorphism, 186 equations, 97

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

347

348 equidivisibility, 327 equivalence of categories, 164 equivalence relation, 1 Esakia space, 143 essentially surjective, 160 faithful, 160 filter, 52 characteristic function of, 58 generated, 60 maximal, 78 prime, 52 filtering, 13 filtering collection, 13 filters partial order on, 53 final state, 275 finite A-word, 275 finite element, 235 finite intersection property, 32 finite model property, 145 finite-join-preserving, 10 finite-meet-preserving, 10 finitely generated distributive lattice, 95 finitely generated down-set, 19 finitely generated up-set, 19, 247, 256 first isomorphism theorem, 12, 282 forgetful functor, 161, 167 forward image, see under functor frame, 12, 201 frame homomorphism, 12, 201 free Boolean algebra, 94 free Boolean extension, 14 free distributive lattice, 87, 167, 193 free join semilattice, 26 free join-semilattice, 168 free monoid, 7, 168, 275 free profinite monoid, 295, 296 free ultrafilter, 81, 100 full infinite binary tree, 4 full subcategory, 159 functional composition, 3 functor, 160 contravariant, 161 covariant, 161 forgetful, 161, 167 full, 160 Gödel translation, 148 Gödel–Löb logic, 126 Galois connection, 6 generalized Cantor space, 92 generated down-set, 19 submonoid, 283 subsemigroup, 283 up-set, 19 graph, 36

Index graph of a binary operation, 302 greatest lower bound, 4 Green’s equivalence relations, 318, 323 H -trivial, 318 Hausdorff maximality principle, 33 Heyting algebra, 132, 139 Heyting homomorphism, 12, 139 Hoffmann–Lawson duality, 238 Hom-class, 155 homeomorphism, 30 homomorphic image, 11 homomorphism between Boolean algebras, 14 between lattices, 10 dual function, 70, 121 dual relations, 121 of monoids, 282 hypercompact, 247 hyperdoctrine, 327 hyperspectral spaces, 246 ideal, 55 generated, 61 prime, 55 idempotent element, 301 operation, 9 relation, 236 identity element, 15, 275 implication type compatible relation of, 133, 302 implication-type operator, 130 inclusion order, 18 incomparable, 1 indiscrete topology, 30 infimum, 4 inherited order, 3 initial object, 177 initial state, 275 interior, 30 interpolating, 236 interval topology, 40 intuitionistic formula, 145 intuitionistic Kripke model, 145 inverse image, 21 isomorphism, 157 between categories, 170 between lattices, 10 isomorphism between ordered spaces, 42 J -trivial, 314, 318 language, 319 join, 4 Join Infinite Distributive law, 12 join irreducible, 19 join preserving at primes, 222, 265 join prime, 20, 222, 252

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

Index join of topologies, 38 join-semilattice, 9 kernel, 11, 282 Kripke Boolean model, 123 Kripke Boolean space, 123 Kripke frame, 123 Kripke model, 123 L-trivial, 318 Löb’s axiom, 126 language recognized by an automaton, 276 lattice algebraic definition, 9 distributive, 12 first isomorphism theorem, 12 order from operations, 9 order-theoretic definition, 9 lattice embedding, 10 lattice homomorphism, 10 lattice isomorphism, 10 lattice term, 89 Lawson topology, 40 least fixed point, 268 least upper bound, 4 left action, 283 lexicographic order, 4, 7 limit in a category, 179 Lindenbaum–Tarski algebra, 90 linear order, 1 locally compact, 32 locally finite, 95 lower bound, 4 lower topology, 40 M3 , 13 machine, 275 magma, 302 map, 2 max, 4 maximal consistent set, 129 maximal vs maximum, 5 meet, 4 meet irreducible, 19 meet prime, 20 meet-semilattice, 9, 193 meet-semilattice reduct, 194 membership problem, 313 min, 4 minimal automaton, 290 minimal upper bounds, 253 modal algebra, 122 positive, 121, 193 modal formula, 121 modal logic classical, 121 positive, 122 modal operator, 121

monoid, 275 congruence, 282 first isomorphism theorem, 282 internal to a category, 176 on a monoid, 282 quotient, 282 monoid compatible, 312 monoid homomorphism, 168 monomorphism, 174 monotone, 2 morphism between ordered spaces, 42 MUB-closed, 261 MUB-closure, 261 MUB-complete, 255 finitely, 255 MV-algebra, 132 N5 , 13 necessity operator, 121 neighborhood, 30 neighborhood filter, 201 neutral dual space, 65 neutral element, 275 one-point compactification, 63, 81 open cover, 32 open mapping, 30 open set, 29 operator of diamond type, 130 of implication type, 141 opposite category, 159 opposite poset, 3 order embedding surjective, 7 order homeomorphism, 42 order ideal, 236 order isomorphism, 2 order normality, 63 order preserving, 2 and injective, 7 order reflecting, 2 order reversing, 3 order-separation property, 43 ordered generalized Cantor space, 49 ordered space, 42 morphism, 42 p-morphism, 144 partial order, 1 patch topology, 45 piecewise testable, 314 Plotkin order, 262 point-closed, 114 points, 66 polynomial closure, 314 poset, 1 poset reflection, 3, 8, 256

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

349

350 positive -modal algebra, 193, 195 positively piecewise testable, 314 power set, 18 prefix order, 4 preorder, 1 preserve directed joins, 230 preserve join at primes, 222 modulo a congruence, 222 for a residuation algebra, 304 Priestley compatible relation downward, 118 Priestley compatible relation upward, 114 Priestley duality, 73 Priestley space, 63 is order-normal, 63 prime filter, 52 characteristic function of, 56 complement, 56 prime ideal, 55 characteristic function of, 56 complement, 56 principal down-set, 19 principal up-set, 19 product, 10 in categories, 175 special case of limit, 179 product space, 31 profinite monoid, 182, 295 profinite poset, 181 profinite set, 182 proper, 50 proper map, 50 pseudocomplement, 139 pseudovariety of finite monoids, 313 pseudovariety of regular languages, 313 pullback, 183 pushout, 183 quasi-identity, 317 quotient of a lattice, 11 of a monoid, 282 of a Priestley space, 102 of a topological space, 37 quotienting operations, 278, 280 R-trivial, 318 reachability relation, 276 recognizer, 287 reduct semilattice, 9 regular epimorphism, 174 regular language, 276 generates finite residuation ideal, 288 has finite syntactic monoid, 288 regular monomorphism, 175

Index rejects, 275 relational composition, 3, 112, 158 relational converse, see under converse relation relational direct image, 113, 278 relational inverse image, 113, 278 relational universal image, 113 relative pseudocomplement, 139 remainder, 81, 100 residual, 131 residuated Boolean algebra dual to a monoid, 280 residuated operation, 131 residuation algebra dual to a monoid, 280 residuation ideal, 280 corresponding to a monoid quotient, 287 generated, 280 right action, 283 Rudin’s lemma, 249 satisfies the (spatial) equation, 105 satisfies the (spatial) inequation, 103 satisfies the profinite monoid equation, 312 satisfies the profinite monoid inequation, 312 saturated, 37 Scott topology, 39, 227, 231 self-dual property, 12 semilattice, 9 semilattice homomorphism, 10 separation axioms, 31 shuffle ideal, 314 Sierpinski space, 45 Simon’s theorem, 293 skeleton, 157, 171 Smyth powerdomain, 194, 212 sober space, 198, 207 dcpo of a, 233 span, 183 spatial, 206 specialization order, 37 spectral dcpo, 246 spectral domain, 228, 246, 251 spectral map, 198 spectral space, 198 spectral subspace embedding, 258 stably compact space, 45 states of an automaton, 275 Stone dual space of a lattice, 198 Stone spaces, 198 Stone’s prime filter-ideal theorem, 57 Stone–Čech compactification, 80 strong unit, 132 subalgebra of a Boolean algebra, 14 subcategory, 159 sublattice, 11 submonoid, 283 subsemigroup, 283

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

Index subspace, 30 subword, 314 supremum, 4 vs maximum, 5 surjective co-restriction, 287 symmetric difference, 17, 110, 177 syntactic congruence, 288 homomorphism, 288 monoid, 288 preorder, 310 term, 89 terminal object, 175 thin, 156 TOD, see under totally order-disconnected tooth, 247 top, 4 topological monoid, 176, 295 topological space, 29 topology generated by a subbase, 31 total order, 1 totally order-disconnected, 63 transition relation, 275 triangle identities, 166 truth set, 123 ultrafilter, 78, 79 free, 81, 100

unbounded sublattice, 11 underlying machine of an automaton, 275 unit of an adjunction, 166 up-directed, 13 up-set, 18 generated by a subset, 19 principal, 19 upper bound, 4 upper Priestley continuous, 114 upper topology, 39 upper Vietoris space, 212, 213 valid formula on a Kripke Boolean model, 124 formula on a Kripke model, 124 valuation, 123 admissible, for intuitionistic Kripke models, 145 admissible, for Kripke Boolean models, 123 Vietoris space, 194 way below, 235 well filtered, 44 well pointed, 171 well quasi-order, 315 Zorn’s lemma, 33

https://doi.org/10.1017/9781009349680.011 Published online by Cambridge University Press

351